Leagues

Leagues

Results f24d1f41, updated 16 November 2026, 21:24 UTC; scored by GNU Backgammon 1.08, 2-ply.

Starting groups

Decision starting group, the original League 1

Swipe for more columns →

Decision: the original League 1's own matches, 171 pairings; 3 points a win, 1 a draw
Model and settingPlaceMatch pointsPlayedWon-drawn-lostScore for-againstCup place · league moves
GNUbg weakprogram, a reference row: an official place in the match standings, not a model type1481816-0-2101-36
d1-ultra2481816-0-290-43Hype Cup seed 5Nackgammon knockout↑ now in League 2round 1: League 1 → 2 (promoted)
GLiDEAPI3371812-1-577-49Hype Cup seed 6Nackgammon knockout↑ now in League 2round 2: League 1 → 2 (promoted)
Jev 1.134361812-0-681-47
Jev 1.145361812-0-677-52↑ now in League 2round 3: League 1 → 2 (promoted)
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell6361812-0-684-46
Cygnet7341811-1-675-59
Clef8321810-2-669-53↑ now in League 2round 4: League 1 → 2 (promoted) · round 5: League 2 → 1 (relegated) · round 6: League 1 → 2 (final swap (ladder ended))
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell9301810-0-869-55
celeris-1-decision1030189-3-674-63↕ back in League 1round 5: League 1 → 2 (promoted) · round 6: League 2 → 1 (final swap (ladder ended))
SubQ Snap Preview1128189-1-876-58
GPT-6 LunaDecisions API1227189-0-962-65
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1327189-0-969-60
Mercury Decide1419186-1-1151-78
Winnow-12B1512184-0-1451-77
Safestprogram, a reference row: an official place in the match standings, not a model type1612183-3-1250-83
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell178182-2-1438-96
Perplexity Decider Flash v1.2186182-0-1636-97
Randomprogram, a reference row: an official place in the match standings, not a model type190180-0-186-119

Level on points, ranked by the tiebreak:

  • 48 points: GNUbg weak, d1-ultra, in that order on head-to-head: d1-ultra 5–6 GNUbg weak (match #192).
  • 36 points: Jev 1.13, Jev 1.14, Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell, in that order on the mini-table of their matches: Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell 0–7 Jev 1.13 (match #416); Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell 3–4 Jev 1.14 (match #415); Jev 1.13 5–1 Jev 1.14 (match #287).
  • 30 points: GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell, celeris-1-decision, in that order on head-to-head: celeris-1-decision 1–5 GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell (match #493).
  • 27 points: GPT-6 LunaDecisions API, Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark, in that order on head-to-head: Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark 2–4 GPT-6 LunaDecisions API (match #254).
  • 12 points: Winnow-12B, Safest, in that order on head-to-head: Winnow-12B 5–3 Safest (match #238).

Workhorse starting group, the original League 2

Swipe for more columns →

Workhorse: the original League 2's own matches, 171 pairings; 3 points a win, 1 a draw
Model and settingPlaceMatch pointsPlayedWon-drawn-lostScore for-againstCup place · league moves
GPT-6.1 Sollow1481816-0-292-37Hype Cup seed 3Nackgammon knockout↑ now in League 3round 1: League 2 → 3 (promoted)
GPT-6.1 Solhigh2421814-0-480-41Hype Cup seed 4Nackgammon knockout↑ now in League 3round 2: League 2 → 3 (promoted)
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark3391813-0-579-54Hype Cup seed 8Nackgammon knockout
Claude Sonnet 5.5high4391813-0-582-46
Gemini 3.8 Flashlow5381812-2-484-47↕ back in League 2round 3: League 2 → 3 (promoted) · round 4: League 3 → 2 (relegated) · round 5: League 2 → 3 (promoted) · round 6: League 3 → 2 (final swap (ladder ended))
Claude Sonnet 5.5low6371812-1-582-48
DeepSeek-V4.1-Flashlow · API · 10177371812-1-588-54
Gemini 3.8 Flashhigh8361812-0-677-48
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release9331811-0-777-55
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1027189-0-958-61
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1125188-1-963-59
GNUbg weakprogram, a reference row: an official place in the match standings, not a model type1224188-0-1067-66
GLM-5.3low1321187-0-1156-78
DeepSeek-V4.1-Flashoff · API · 10171415185-0-1356-82
Qwen3.8 Flashoff1515185-0-1350-86
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release1613184-1-1356-83↓ now in League 1round 4: League 2 → 1 (relegated)
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1712184-0-1437-91↓ now in League 1round 3: League 2 → 1 (relegated)
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark189183-0-1539-98↓ now in League 1round 2: League 2 → 1 (relegated)
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell190180-0-1813-102↓ now in League 1round 1: League 2 → 1 (relegated)

Level on points, ranked by the tiebreak:

  • 39 points: Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark, Claude Sonnet 5.5high, in that order on head-to-head: Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark 4–2 Claude Sonnet 5.5high (match #202).
  • 37 points: Claude Sonnet 5.5low, DeepSeek-V4.1-Flashlow · API · 1017, in that order on head-to-head: Claude Sonnet 5.5low 7–4 DeepSeek-V4.1-Flashlow · API · 1017 (match #613).
  • 15 points: DeepSeek-V4.1-Flashoff · API · 1017, Qwen3.8 Flashoff, in that order on head-to-head: DeepSeek-V4.1-Flashoff · API · 1017 6–1 Qwen3.8 Flashoff (match #629).

Frontier starting group, the original League 3

Swipe for more columns →

Frontier: the original League 3's own matches, 66 pairings; 3 points a win, 1 a draw
Model and settingPlaceMatch pointsPlayedWon-drawn-lostScore for-againstCup place · league moves
GPT-6 Astramedium127119-0-249-30Hype Cup seed 1Nackgammon knockout
Claude Fable 5.1high224118-0-351-31Hype Cup seed 2Nackgammon knockout
Grok 4.7high319116-1-439-40Hype Cup seed 7Nackgammon knockout
Claude Opus 5.5high418116-0-548-37
GPT-6 Astralow517115-2-444-33
Gemini 4 Argonhigh616115-1-536-43
Gemini 4 Argonlow716115-1-538-45↕ back in League 3round 3: League 3 → 2 (relegated) · round 4: League 2 → 3 (promoted) · round 5: League 3 → 2 (relegated) · round 6: League 2 → 3 (final swap (ladder ended))
Claude Opus 5.5low815115-0-638-42
Kimi K3.1 Prooff915115-0-638-38
Grok 4.7low1012114-0-733-40↓ now in League 2round 2: League 3 → 2 (relegated)
GNUbg weakprogram, a reference row: an official place in the match standings, not a model type118112-2-732-54
Claude Fable 5.1low127112-1-837-50↓ now in League 2round 1: League 3 → 2 (relegated)

Level on points, ranked by the tiebreak:

  • 16 points: Gemini 4 Argonhigh, Gemini 4 Argonlow, in that order on head-to-head: Gemini 4 Argonlow 2–4 Gemini 4 Argonhigh (match #432).
  • 15 points: Claude Opus 5.5low, Kimi K3.1 Prooff, in that order on head-to-head: Kimi K3.1 Prooff 4–5 Claude Opus 5.5low (match #603).

Groups reflect model profiles, not tournament rankings.

These are fixed starting cohorts, not current ladder ranks or mutually exclusive universal model types.

Each setup stays in the group it started in, the original league it played in before any league move, with that group's own match points: only matches between setups of the same starting group count, wherever they were played (league season, ladder round or gap match); late entries are left out. 3 points a win, 1 a draw. Ties: points, then head-to-head (a mini-table of the matches among three or more), then the lower error rate (rule set on 28 Oct, after the ties were known; no place here needed the error rate). GNUbg weak, the anchor program, is in all three groups. A setup that moved between leagues keeps its row here; where it plays now is in After league moves.

Badges: Hype Cup seed a Cup place (Cups) · ↑ now in League N ↓ now in League N ↕ back in League N league moves, from the ladder's records. Match points are match results only; they are not the Risk, Quality or Confidence measures.

After league moves

League 1

FinishedPublished

Swipe for more columns →

Standings: 3 points a win, 1 a draw; only finished league-stage matches count
Model and settingPlacePlayedWon-drawn-lostPointsGames won-lostGammons wonTo playStarting groupLeague moves
GNUbg weakprogram · CPU11816-0-24854-22350Decision, Workhorse, Frontier program, original Leagues 1, 2, 3
celeris-1-decisiondecision model · API21812-2-43857-39170Decision original League 1↕ back in League 1round 5: League 1 → 2 (promoted) · round 6: League 2 → 1 (final swap (ladder ended))
SubQ Snap Previewdecision model · API31812-1-53766-38140Decision original League 1
CygnetNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs41812-1-53760-40150Decision original League 1
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs51812-0-63660-36230Decision original League 1
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs61812-0-63659-39180Decision original League 1
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release71812-0-63658-28170Workhorse original League 2↓ now in League 1round 4: League 2 → 1 (relegated)
Jev 1.13decision model · API81812-0-63655-39260Decision original League 1
GPT-6 LunaDecisions APIdecision model · API91811-1-63451-47110Decision original League 1
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs101810-0-83054-48230Decision original League 1
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark111810-0-83050-50150Workhorse original League 2↓ now in League 1round 2: League 2 → 1 (relegated)
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark12189-1-82851-5390Workhorse original League 2↓ now in League 1round 3: League 2 → 1 (relegated)
Mercury Decidedecision model · API13188-1-92543-53100Decision original League 1
Winnow-12BNVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs14187-0-112147-53140Decision original League 1
Safestprogram · CPU15184-2-121441-65100Decision program, original League 1
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell16183-0-15931-63110Workhorse original League 2↓ now in League 1round 1: League 2 → 1 (relegated)
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs17182-1-15730-7050Decision original League 1
Perplexity Decider Flash v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs18182-0-16633-6570Decision original League 1
Randomprogram · CPU19180-0-1806-5800Decision program, original League 1

171 of 171 matches finished, 906 games; started 2026-10-15, finished 2026-10-26. 36,995 of 36,995 decisions engine-scored so far.

Notes

  • Won: GNUbg weak, 48 points from 16-0-2 in 18 matches.
  • Lowest error rate in league play: Jev 1.13, 34.4 (31.0 to 38.0); no clear difference in league play from GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell, DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release, celeris-1-decision, Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell, Cygnet, Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark.
  • LFM2.5-24B-A2B, off → on: lower error rate in league play.
  • League comparisons are not paired: each setup met different positions. Paired comparisons come from the shared positions.
  • The anchor is GNU Backgammon 1.08 Beginner, x86-64 build; runner changes during the leagues are listed in Method.

Decision quality

Swipe for more columns →

Decision quality in this league's games; intervals resample whole games
Model and settingError rate (PR), 95% intervalBand on the human scaleEngine's best playBlunder rateMedian / P90 s per decision$ per 1,000 decisionsDecisions scoredInvalid answers
GNUbg weakprogram · CPU21.418.8 to 24.3Casual (approx.)52.0%19.0%0.01 / 0.01Not measured1,7440
Jev 1.13decision model · API34.431.0 to 38.0Distracted (approx.)50.7%25.1%0.22 / 0.260.166API price1,9110
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs35.031.6 to 39.0Distracted (approx.)47.1%24.6%1.41 / 2.7510.301estimate · RTX PRO 6000 Blackwell2,003†1 split into groups0
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release35.631.7 to 39.7Distracted (approx.)47.9%25.2%0.48 / 1.630.728estimate · RTX PRO 6000 Blackwell + API1,736†147 split into groups0
celeris-1-decisiondecision model · API35.632.0 to 39.4Distracted (approx.)45.8%25.8%0.17 / 0.330.150API price1,9220
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs36.933.9 to 40.1Distracted (approx.)46.6%27.9%0.94 / 2.7312.227estimate · RTX PRO 6000 Blackwell2,028†206 split into groups0
CygnetNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs38.635.4 to 41.9Distracted (approx.)44.0%27.5%0.37 / 2.9611.168estimate · RTX PRO 6000 Blackwell2,016†123 split into groups0
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs38.835.3 to 42.5Distracted (approx.)47.1%26.8%1.97 / 7.486.593estimate · RTX PRO 6000 Blackwell2,052†226 split into groups0
GPT-6 LunaDecisions APIdecision model · API43.339.7 to 46.9Distracted (approx.)43.2%28.9%0.23 / 0.350.335API price1,9620
SubQ Snap Previewdecision model · API43.939.3 to 48.8Distracted (approx.)42.3%31.8%0.24 / 0.320.193API price2,200†1 split into groups0
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark45.040.7 to 49.9Distracted (approx.)43.7%30.7%1.48 / 2.714.551estimate · RTX PRO 6000 Blackwell2,005†180 split into groups0
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark47.744.1 to 51.2Distracted (approx.)39.6%32.4%324.8 / 668.638.716estimate · RTX PRO 6000 Blackwell2,193†229 split into groups0
Mercury Decidedecision model · API47.943.3 to 52.9Distracted (approx.)42.2%30.3%0.48 / 0.770.071API price1,8530
Winnow-12BNVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs49.745.9 to 53.6Distracted (approx.)38.9%32.3%2.46 / 5.459.615estimate · RTX PRO 6000 Blackwell2,015†92 split into groups0
Safestprogram · CPU51.647.4 to 56.2Distracted (approx.)37.3%35.1%0.00 / 0.00Not measured2,0580
Perplexity Decider Flash v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs59.654.9 to 64.5Distracted (approx.)37.4%37.5%0.56 / 0.9612.267estimate · RTX PRO 6000 Blackwell2,006†224 split into groups0
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell61.556.3 to 66.8Distracted (approx.)32.2%40.8%0.14 / 0.369.132estimate · RTX PRO 6000 Blackwell1,952†182 split into groups0
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs65.659.7 to 71.9Distracted (approx.)36.6%38.6%1.57 / 3.387.185estimate · RTX PRO 6000 Blackwell2,0020
Randomprogram · CPU121.1113.3 to 128.5Distracted (approx.)16.5%61.5%0.00 / 0.00Not measured1,3370

† Setups got different inputs: when a position had more legal plays than a setup takes in one request, its plays were split into groups and the group winners compared. The rows marked † had such decisions; the others saw every play at once. Every setup got the same long rules text.

Time and cost

Swipe for more columns →

Time and cost in this league
Model and setting$ per 1,000 decisionsMedian / P90 s per decisionMean retry/wait sOver the time budgetTokens in / out per decisionCost coverageDecisions
GNUbg weakprogram · CPUNot measured0.01 / 0.010.000.0%Not measured0%1,744
Jev 1.13decision model · API0.166API price0.22 / 0.260.000.0%3,945 / 257100%1,911
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs10.301estimate · RTX PRO 6000 Blackwell1.41 / 2.750.000.0%3,800 / 0100%2,003
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release0.728estimate · RTX PRO 6000 Blackwell + API0.48 / 1.6315.50.1%3,960 / 9100%1,736
celeris-1-decisiondecision model · API0.150API price0.17 / 0.330.000.0%3,747 / 0100%1,922
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs12.227estimate · RTX PRO 6000 Blackwell0.94 / 2.731.840.0%4,331 / 0100%2,028
CygnetNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs11.168estimate · RTX PRO 6000 Blackwell0.37 / 2.960.000.0%4,593 / 1100%2,016
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs6.593estimate · RTX PRO 6000 Blackwell1.97 / 7.483.860.6%4,488 / 0100%2,052
GPT-6 LunaDecisions APIdecision model · API0.335API price0.23 / 0.350.010.0%3,351 / 0100%1,962
SubQ Snap Previewdecision model · API0.193API price0.24 / 0.320.000.0%3,866 / 0100%2,200
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4.551estimate · RTX PRO 6000 Blackwell1.48 / 2.710.420.0%4,266 / 3100%2,005
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark38.716estimate · RTX PRO 6000 Blackwell324.8 / 668.6127.7100.0%28,836 / 21,668100%2,193
Mercury Decidedecision model · API0.071API price0.48 / 0.770.010.0%3,567 / 1100%1,853
Winnow-12BNVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs9.615estimate · RTX PRO 6000 Blackwell2.46 / 5.450.000.1%4,144 / 0100%2,015
Safestprogram · CPUNot measured0.00 / 0.000.000.0%Not measured0%2,058
Perplexity Decider Flash v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs12.267estimate · RTX PRO 6000 Blackwell0.56 / 0.960.700.0%4,492 / 0100%2,006
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell9.132estimate · RTX PRO 6000 Blackwell0.14 / 0.360.000.0%3,882 / 1100%1,952
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs7.185estimate · RTX PRO 6000 Blackwell1.57 / 3.382.020.0%35,473 / 0100%2,002
Randomprogram · CPUNot measured0.00 / 0.000.000.0%Not measured0%1,337
Matches (171)

Swipe for more columns →

Matches
MatchMatchdayScoreStatusGames
SubQ Snap Preview v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark15–6Finished8
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v GNUbg weak20–7Finished4
Random v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell30–6Finished4
Safest v SubQ Snap Preview40–6Finished6
Cygnet v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell45–1Finished6
GNUbg weak v Safest57–3Finished6
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell57–1Finished4
GNUbg weak v Random65–0Finished2
Safest v Random77–0Finished4
Cygnet v GNUbg weak70–7Finished4
SubQ Snap Preview v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell86–0Finished6
Cygnet v Random85–0Finished2
GNUbg weak v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark85–1Finished4
Random v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark90–7Finished4
Safest v Cygnet94–4Finished8
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Cygnet106–3Finished6
Safest v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark114–3Finished6
SubQ Snap Preview v GNUbg weak111–5Finished4
Safest v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell125–5Finished8
Random v SubQ Snap Preview120–8Finished4
SubQ Snap Preview v Cygnet133–7Finished6
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v GNUbg weak150–5Finished2
Jev 1.13 v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark166–3Finished8
Jev 1.13 v SubQ Snap Preview162–5Finished6
Jev 1.13 v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell166–0Finished4
Cygnet v Jev 1.13162–4Finished6
Random v Jev 1.13160–6Finished2
Jev 1.13 v Safest165–4Finished6
Jev 1.13 v GNUbg weak166–0Finished4
GNUbg weak v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark167–2Finished4
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell168–4Finished6
GNUbg weak v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark175–0Finished4
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell174–3Finished6
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark171–5Finished6
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GPT-6 LunaDecisions API182–4Finished6
GPT-6 LunaDecisions API v SubQ Snap Preview180–7Finished4
GPT-6 LunaDecisions API v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell185–3Finished8
GPT-6 LunaDecisions API v Cygnet185–4Finished6
GPT-6 LunaDecisions API v Random186–1Finished4
Safest v GPT-6 LunaDecisions API180–5Finished4
GPT-6 LunaDecisions API v GNUbg weak181–5Finished4
Jev 1.13 v GPT-6 LunaDecisions API184–5Finished6
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark180–7Finished4
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v SubQ Snap Preview182–4Finished6
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell184–5Finished6
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v Cygnet182–4Finished6
Random v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell181–6Finished4
Safest v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell185–3Finished4
Jev 1.13 v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell187–1Finished6
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v GPT-6 LunaDecisions API180–5Finished4
Perplexity Decider Flash v1.2 v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell182–8Finished6
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v Winnow-12B184–6Finished6
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v Mercury Decide180–6Finished6
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell184–7Finished8
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell185–3Finished6
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark184–5Finished6
SubQ Snap Preview v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark186–1Finished6
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell186–1Finished6
Cygnet v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark184–5Finished6
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Random185–0Finished2
Safest v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark182–4Finished6
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Jev 1.13180–6Finished4
GPT-6 LunaDecisions API v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark184–3Finished6
Perplexity Decider Flash v1.2 v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark183–7Finished6
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Winnow-12B183–6Finished6
Mercury Decide v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark181–5Finished6
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark182–5Finished6
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark186–3Finished6
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark185–8Finished8
SubQ Snap Preview v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark186–2Finished6
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark183–6Finished6
Cygnet v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark186–0Finished6
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Random186–0Finished4
Safest v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark182–5Finished6
Jev 1.13 v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark183–4Finished6
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GPT-6 LunaDecisions API184–4Finished8
Perplexity Decider Flash v1.2 v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark183–6Finished6
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Winnow-12B182–5Finished6
Mercury Decide v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark184–3Finished6
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell185–3Finished8
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell185–2Finished6
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v GNUbg weak181–5Finished4
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release180–6Finished4
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark182–6Finished6
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark188–0Finished4
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Perplexity Decider Flash v1.2195–3Finished6
SubQ Snap Preview v Perplexity Decider Flash v1.2195–1Finished4
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider Flash v1.2195–3Finished8
Cygnet v Perplexity Decider Flash v1.2196–1Finished4
Perplexity Decider Flash v1.2 v Random197–3Finished6
Safest v Perplexity Decider Flash v1.2195–3Finished6
GNUbg weak v Perplexity Decider Flash v1.2199–0Finished4
Jev 1.13 v Perplexity Decider Flash v1.2195–3Finished6
Perplexity Decider Flash v1.2 v GPT-6 LunaDecisions API191–6Finished6
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark196–2Finished4
SubQ Snap Preview v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release192–4Finished6
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release190–5Finished4
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v Cygnet193–4Finished6
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v Random195–0Finished2
Safest v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release190–5Finished4
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v Jev 1.13195–3Finished6
GPT-6 LunaDecisions API v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release190–6Finished4
Perplexity Decider Flash v1.2 v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release191–5Finished4
Winnow-12B v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release191–5Finished6
Mercury Decide v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release194–3Finished6
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell196–0Finished4
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release198–3Finished6
celeris-1-decision v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark193–4Finished6
SubQ Snap Preview v celeris-1-decision194–4Finished8
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v celeris-1-decision191–5Finished4
celeris-1-decision v Cygnet192–5Finished6
celeris-1-decision v Random197–0Finished4
Safest v celeris-1-decision192–4Finished6
GNUbg weak v celeris-1-decision193–5Finished4
Jev 1.13 v celeris-1-decision192–5Finished6
celeris-1-decision v GPT-6 LunaDecisions API196–3Finished6
celeris-1-decision v Perplexity Decider Flash v1.2195–3Finished6
celeris-1-decision v Winnow-12B195–3Finished6
celeris-1-decision v Mercury Decide195–5Finished8
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v celeris-1-decision196–2Finished4
celeris-1-decision v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell191–5Finished4
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell v celeris-1-decision191–5Finished4
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v celeris-1-decision192–4Finished6
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v celeris-1-decision190–5Finished2
celeris-1-decision v DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release195–3Finished6
Winnow-12B v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark201–5Finished4
Winnow-12B v SubQ Snap Preview202–7Finished6
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v Winnow-12B201–5Finished6
Cygnet v Winnow-12B206–2Finished4
Random v Winnow-12B200–7Finished4
Winnow-12B v Safest205–3Finished8
Winnow-12B v GNUbg weak204–5Finished6
Winnow-12B v Jev 1.13202–5Finished4
Winnow-12B v GPT-6 LunaDecisions API204–3Finished6
Perplexity Decider Flash v1.2 v Winnow-12B205–3Finished6
Mercury Decide v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark212–5Finished6
SubQ Snap Preview v Mercury Decide215–2Finished6
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v Mercury Decide213–7Finished6
Mercury Decide v Cygnet214–5Finished6
Mercury Decide v Random216–1Finished4
Mercury Decide v Safest214–3Finished6
Mercury Decide v GNUbg weak210–8Finished4
Jev 1.13 v Mercury Decide216–0Finished4
Mercury Decide v GPT-6 LunaDecisions API212–5Finished6
Perplexity Decider Flash v1.2 v Mercury Decide211–5Finished4
Mercury Decide v Winnow-12B214–3Finished6
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark234–3Finished6
SubQ Snap Preview v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell230–7Finished6
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell230–6Finished4
Cygnet v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell234–2Finished6
Random v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell230–8Finished4
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v Safest234–3Finished6
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v GNUbg weak234–5Finished6
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v Jev 1.13230–7Finished4
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v GPT-6 LunaDecisions API236–0Finished4
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider Flash v1.2238–2Finished6
Winnow-12B v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell230–5Finished4
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell v Mercury Decide237–0Finished4
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark244–3Finished6
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v SubQ Snap Preview242–4Finished6
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell242–5Finished6
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v Cygnet244–5Finished6
Random v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell240–8Finished4
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v Safest245–1Finished6
GNUbg weak v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell247–1Finished6
Jev 1.13 v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell242–5Finished6
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v GPT-6 LunaDecisions API243–4Finished6
Perplexity Decider Flash v1.2 v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell240–5Finished4
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v Winnow-12B245–2Finished6
Mercury Decide v GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell240–5Finished2
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell v Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell242–6Finished6

This league table's matches. Every match of every stage (ladder rounds, gap matches, late entries, the Hype Cup, the Nackgammon knockout) and the voided ones: Matches.

League 2

FinishedPublished

Swipe for more columns →

Standings: 3 points a win, 1 a draw; only finished league-stage matches count
Model and settingPlacePlayedWon-drawn-lostPointsGames won-lostGammons wonTo playStarting groupLeague moves
Gemini 3.8 Flashlowchat model, one letter · subscription11814-2-24463-39220Workhorse original League 2↕ back in League 2round 3: League 2 → 3 (promoted) · round 4: League 3 → 2 (relegated) · round 5: League 2 → 3 (promoted) · round 6: League 3 → 2 (final swap (ladder ended))
Claude Sonnet 5.5lowchat model, one letter · subscription21813-1-44065-41170Workhorse original League 2
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs31813-0-53962-42110Workhorse original League 2
DeepSeek-V4.1-Flashlow · API · 1017chat model, answer-letter probabilities · API41812-1-53762-46230Workhorse original League 2
Claude Sonnet 5.5highchat model, one letter · subscription51812-0-63661-41110Workhorse original League 2
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first releasechat model, answer-letter probabilities · API61812-0-63659-43150Workhorse original League 2
Gemini 3.8 Flashhighchat model, one letter · subscription71812-0-63657-47110Workhorse original League 2
Grok 4.7low81811-0-73359-43140Frontier original League 3↓ now in League 2round 2: League 3 → 2 (relegated)
Claude Fable 5.1low91810-0-83056-42130Frontier original League 3↓ now in League 2round 1: League 3 → 2 (relegated)
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs101810-0-83052-42140Workhorse original League 2
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs11189-0-92753-4560Workhorse original League 2
GNUbg weakprogram · CPU12188-0-102442-52210Decision, Workhorse, Frontier program, original Leagues 1, 2, 3
GLM-5.3lowchat model, answer-letter probabilities · API13187-0-112139-61120Workhorse original League 2
d1-ultra14186-1-111946-56100Decision original League 1↑ now in League 2round 1: League 1 → 2 (promoted)
DeepSeek-V4.1-Flashoff · API · 1017chat model, answer-letter probabilities · API15185-1-121639-63130Workhorse original League 2
GLiDEAPI16185-0-131545-5990Decision original League 1↑ now in League 2round 2: League 1 → 2 (promoted)
Jev 1.1417185-0-131534-5880Decision original League 1↑ now in League 2round 3: League 1 → 2 (promoted)
Qwen3.8 Flashoffchat model, answer-letter probabilities · API18182-0-16633-6780Workhorse original League 2
ClefNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark19182-0-16627-6780Decision original League 1↑ now in League 2round 4: League 1 → 2 (promoted) · round 5: League 2 → 1 (relegated) · round 6: League 1 → 2 (final swap (ladder ended))

171 of 171 matches finished, 954 games; started 2026-10-19, finished 2026-10-26. 39,538 of 39,538 decisions engine-scored so far.

Notes

  • Won: Gemini 3.8 Flashlow, 44 points from 14-2-2 in 18 matches.
  • Lowest error rate in league play: Gemini 3.8 Flashhigh, 14.9 (12.9 to 17.0); no clear difference in league play from Gemini 3.8 Flashlow, Claude Sonnet 5.5high, DeepSeek-V4.1-Flashlow · API · 1017, Claude Fable 5.1low, DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release, Grok 4.7low.
  • Claude Sonnet 5.5, low → high: no clear difference in league play.
  • DeepSeek-V4.1-Flash1017, off → low: lower error rate in league play.
  • Gemini 3.8 Flash, low → high: no clear difference in league play.
  • Gemma 4 31B, off → on: no clear difference in league play.
  • League comparisons are not paired: each setup met different positions. Paired comparisons come from the shared positions.
  • The anchor is GNU Backgammon 1.08 Beginner, x86-64 build; runner changes during the leagues are listed in Method.

Decision quality

Swipe for more columns →

Decision quality in this league's games; intervals resample whole games
Model and settingError rate (PR), 95% intervalBand on the human scaleEngine's best playBlunder rateMedian / P90 s per decision$ per 1,000 decisionsDecisions scoredInvalid answers
Gemini 3.8 Flashhighchat model, one letter · subscription14.912.9 to 17.0Intermediate (approx.)63.1%11.0%9.59 / 44.422.352API-equivalent (subscription)2,168†237 split into groups0
Gemini 3.8 Flashlowchat model, one letter · subscription16.314.5 to 18.2Intermediate (approx.)59.0%13.5%2.16 / 6.874.123API-equivalent (subscription)2,053†202 split into groups0
Claude Sonnet 5.5highchat model, one letter · subscription16.714.9 to 18.7Intermediate (approx.)62.5%13.3%6.63 / 17.619.847API-equivalent (subscription)2,164†257 split into groups0
DeepSeek-V4.1-Flashlow · API · 1017chat model, answer-letter probabilities · API17.115.1 to 19.2Intermediate (approx.)61.8%13.4%22.3 / 107.64.869API price2,222†258 split into groups12
Claude Fable 5.1low17.915.9 to 20.1Casual (approx.)58.7%15.1%1.54 / 6.3462.613API-equivalent (subscription)2,008†232 split into groups0
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first releasechat model, answer-letter probabilities · API19.016.6 to 21.3Casual (approx.)59.7%13.9%43.8 / 173.147.092estimate · RTX PRO 6000 Blackwell + API2,188†295 split into groups38
Grok 4.7low19.116.9 to 21.4Casual (approx.)60.8%15.1%24.3 / 110.232.813API-equivalent (subscription)2,152†250 split into groups0
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs19.617.5 to 21.7Casual (approx.)59.1%15.6%120.6 / 394.334.943estimate · RTX PRO 6000 Blackwell2,122†252 split into groups0
Claude Sonnet 5.5lowchat model, one letter · subscription20.117.8 to 22.4Casual (approx.)60.1%15.7%1.51 / 7.3713.541API-equivalent (subscription)2,150†217 split into groups0
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs22.519.9 to 25.3Casual (approx.)54.4%18.2%0.31 / 4.038.215estimate · RTX PRO 6000 Blackwell1,847†183 split into groups0
GNUbg weakprogram · CPU27.024.1 to 29.8Beginner (approx.)46.2%22.1%0.01 / 0.01Not measured2,1360
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs29.026.5 to 31.8Beginner (approx.)52.8%22.4%171.1 / 566.739.023estimate · RTX PRO 6000 Blackwell + API2,108†217 split into groups0
d1-ultra30.327.3 to 33.4Distracted (approx.)48.6%22.4%0.23 / 0.410.145API price2,071†1 split into groups0
GLM-5.3lowchat model, answer-letter probabilities · API30.927.8 to 34.0Distracted (approx.)50.2%21.7%1.75 / 7.482.917API price1,989†203 split into groups0
Jev 1.1431.728.3 to 35.7Distracted (approx.)46.2%24.7%0.19 / 0.240.173API price1,912†2 split into groups0
GLiDEAPI33.230.2 to 36.4Distracted (approx.)48.4%24.1%0.48 / 0.790.153API price2,159†2 split into groups0
Qwen3.8 Flashoffchat model, answer-letter probabilities · API33.630.8 to 36.7Distracted (approx.)44.6%25.3%1.70 / 7.470.284API price2,027†171 split into groups0
DeepSeek-V4.1-Flashoff · API · 1017chat model, answer-letter probabilities · API35.231.8 to 38.8Distracted (approx.)47.7%23.6%0.84 / 1.390.244API price2,034†169 split into groups0
ClefNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark36.533.5 to 39.6Distracted (approx.)44.5%27.1%1.13 / 5.415.736estimate · RTX PRO 6000 Blackwell1,978†22 split into groups0

† Setups got different inputs: when a position had more legal plays than a setup takes in one request, its plays were split into groups and the group winners compared. The rows marked † had such decisions; the others saw every play at once. Every setup got the same long rules text.

Time and cost

Swipe for more columns →

Time and cost in this league
Model and setting$ per 1,000 decisionsMedian / P90 s per decisionMean retry/wait sOver the time budgetTokens in / out per decisionCost coverageDecisions
Gemini 3.8 Flashhighchat model, one letter · subscription22.352API-equivalent (subscription)9.59 / 44.42.2918.4%4,344 / 5,092100%2,168
Gemini 3.8 Flashlowchat model, one letter · subscription4.123API-equivalent (subscription)2.16 / 6.870.040.7%4,311 / 237100%2,053
Claude Sonnet 5.5highchat model, one letter · subscription19.847API-equivalent (subscription)6.63 / 17.60.263.4%5,664 / 852100%2,164
DeepSeek-V4.1-Flashlow · API · 1017chat model, answer-letter probabilities · API4.869API price22.3 / 107.612.143.4%4,215 / 7,173100%2,234
Claude Fable 5.1low62.613API-equivalent (subscription)1.54 / 6.340.041.1%5,728 / 107100%2,008
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first releasechat model, answer-letter probabilities · API47.092estimate · RTX PRO 6000 Blackwell + API43.8 / 173.11.9963.7%4,364 / 12,487100%2,226
Grok 4.7low32.813API-equivalent (subscription)24.3 / 110.214.042.7%5,999 / 3,469100%2,152
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs34.943estimate · RTX PRO 6000 Blackwell120.6 / 394.315.396.4%14,746 / 5,599100%2,122
Claude Sonnet 5.5lowchat model, one letter · subscription13.541API-equivalent (subscription)1.51 / 7.370.010.5%5,579 / 238100%2,150
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs8.215estimate · RTX PRO 6000 Blackwell0.31 / 4.030.901.6%4,167 / 1100%1,847
GNUbg weakprogram · CPUNot measured0.01 / 0.010.000.0%Not measured0%2,136
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkchat model, answer-letter probabilities · our GPUs39.023estimate · RTX PRO 6000 Blackwell + API171.1 / 566.724.097.7%14,723 / 5,772100%2,108
d1-ultra0.145API price0.23 / 0.410.000.0%3,461 / 0100%2,071
GLM-5.3lowchat model, answer-letter probabilities · API2.917API price1.75 / 7.486.690.6%4,005 / 148100%1,989
Jev 1.140.173API price0.19 / 0.240.000.0%4,116 / 292100%1,912
GLiDEAPI0.153API price0.48 / 0.790.000.0%3,829 / 0100%2,159
Qwen3.8 Flashoffchat model, answer-letter probabilities · API0.284API price1.70 / 7.478.970.3%4,182 / 10100%2,027
DeepSeek-V4.1-Flashoff · API · 1017chat model, answer-letter probabilities · API0.244API price0.84 / 1.3924.20.0%3,846 / 1100%2,034
ClefNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark5.736estimate · RTX PRO 6000 Blackwell1.13 / 5.412.150.2%4,028 / 0100%1,978
Matches (171)

Swipe for more columns →

Matches
MatchMatchdayScoreStatusGames
Gemini 3.8 Flashhigh v GLM-5.3low15–3Finished6
Gemini 3.8 Flashhigh v Claude Sonnet 5.5low25–4Finished8
GNUbg weak v Gemini 3.8 Flashhigh35–0Finished4
Claude Sonnet 5.5high v GLM-5.3low35–3Finished8
Gemini 3.8 Flashhigh v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release43–4Finished6
Claude Sonnet 5.5high v Claude Sonnet 5.5low44–2Finished6
GLM-5.3low v Claude Sonnet 5.5low50–5Finished4
GNUbg weak v Claude Sonnet 5.5high50–8Finished6
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v Claude Sonnet 5.5high65–1Finished4
GLM-5.3low v GNUbg weak62–5Finished4
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v GLM-5.3low73–4Finished6
GNUbg weak v Claude Sonnet 5.5low73–4Finished6
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v Claude Sonnet 5.5low81–7Finished6
Gemini 3.8 Flashhigh v Claude Sonnet 5.5high94–6Finished8
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v GNUbg weak94–3Finished6
Qwen3.8 Flashoff v Gemini 3.8 Flashhigh102–5Finished6
Claude Sonnet 5.5low v Qwen3.8 Flashoff106–1Finished6
Claude Sonnet 5.5high v Qwen3.8 Flashoff104–3Finished6
Qwen3.8 Flashoff v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release103–6Finished6
Qwen3.8 Flashoff v GLM-5.3low103–4Finished6
Qwen3.8 Flashoff v GNUbg weak105–3Finished6
d1-ultra v GNUbg weak105–6Finished6
Claude Fable 5.1low v GNUbg weak113–6Finished6
GNUbg weak v GLiDEAPI125–1Finished4
d1-ultra v GLiDEAPI125–3Finished6
Gemini 3.8 Flashhigh v Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark132–5Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Claude Sonnet 5.5low133–4Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Claude Sonnet 5.5high134–2Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release131–7Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GLM-5.3low134–2Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GNUbg weak136–4Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Qwen3.8 Flashoff137–1Finished6
GNUbg weak v Grok 4.7low132–5Finished6
Claude Fable 5.1low v Grok 4.7low132–4Finished6
Gemini 3.8 Flashhigh v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark144–2Finished6
Claude Sonnet 5.5low v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark142–5Finished4
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Claude Sonnet 5.5high142–4Finished6
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark145–2Finished4
GLM-5.3low v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark143–5Finished4
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GNUbg weak142–5Finished4
Qwen3.8 Flashoff v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark141–6Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark145–2Finished6
Gemini 3.8 Flashhigh v GLiDEAPI151–6Finished6
Claude Sonnet 5.5low v GLiDEAPI155–3Finished6
Claude Sonnet 5.5high v GLiDEAPI155–2Finished4
GLiDEAPI v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release152–5Finished6
GLiDEAPI v GLM-5.3low153–4Finished6
GLiDEAPI v Qwen3.8 Flashoff154–3Finished6
GLiDEAPI v Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark152–4Finished6
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GLiDEAPI154–2Finished6
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GLiDEAPI156–1Finished6
Claude Fable 5.1low v GLiDEAPI156–0Finished4
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Gemini 3.8 Flashhigh161–5Finished6
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Claude Sonnet 5.5low162–5Finished6
Claude Sonnet 5.5high v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark163–4Finished6
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark165–1Finished4
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GLM-5.3low165–1Finished4
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v GNUbg weak165–1Finished4
Qwen3.8 Flashoff v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark165–3Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark165–0Finished4
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark162–4Finished6
Gemini 3.8 Flashhigh v d1-ultra165–1Finished4
Claude Sonnet 5.5low v d1-ultra165–1Finished4
d1-ultra v Claude Sonnet 5.5high163–4Finished6
d1-ultra v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release162–8Finished6
GLM-5.3low v d1-ultra165–4Finished6
Qwen3.8 Flashoff v d1-ultra162–4Finished6
d1-ultra v Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark162–4Finished6
d1-ultra v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark162–5Finished6
d1-ultra v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark164–3Finished6
Gemini 3.8 Flashhigh v Grok 4.7low165–1Finished4
Grok 4.7low v Claude Sonnet 5.5low162–5Finished6
Grok 4.7low v Claude Sonnet 5.5high167–2Finished6
Grok 4.7low v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release165–3Finished8
Grok 4.7low v GLM-5.3low164–5Finished6
Qwen3.8 Flashoff v Grok 4.7low161–5Finished4
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Grok 4.7low162–4Finished6
Grok 4.7low v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark165–1Finished4
Grok 4.7low v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark163–4Finished6
Grok 4.7low v d1-ultra163–4Finished6
GLiDEAPI v Grok 4.7low163–5Finished8
GNUbg weak v Jev 1.14167–2Finished4
d1-ultra v Jev 1.14164–2Finished6
GLiDEAPI v Jev 1.14165–3Finished6
Claude Fable 5.1low v Gemini 3.8 Flashhigh174–5Finished6
Claude Sonnet 5.5low v Claude Fable 5.1low171–5Finished6
Claude Fable 5.1low v Claude Sonnet 5.5high175–1Finished6
Claude Fable 5.1low v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release171–5Finished4
GLM-5.3low v Claude Fable 5.1low174–3Finished6
Qwen3.8 Flashoff v Claude Fable 5.1low172–7Finished6
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Claude Fable 5.1low173–5Finished6
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Claude Fable 5.1low176–1Finished6
Claude Fable 5.1low v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark174–2Finished6
d1-ultra v Claude Fable 5.1low171–5Finished4
Gemini 3.8 Flashhigh v DeepSeek-V4.1-Flashoff · API · 1017175–0Finished4
Claude Sonnet 5.5low v DeepSeek-V4.1-Flashoff · API · 1017175–3Finished6
Claude Sonnet 5.5high v DeepSeek-V4.1-Flashoff · API · 1017176–0Finished4
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v DeepSeek-V4.1-Flashoff · API · 1017172–4Finished6
DeepSeek-V4.1-Flashoff · API · 1017 v GLM-5.3low176–7Finished8
DeepSeek-V4.1-Flashoff · API · 1017 v GNUbg weak176–5Finished6
DeepSeek-V4.1-Flashoff · API · 1017 v Qwen3.8 Flashoff176–1Finished4
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v DeepSeek-V4.1-Flashoff · API · 1017175–3Finished6
DeepSeek-V4.1-Flashoff · API · 1017 v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark172–5Finished6
DeepSeek-V4.1-Flashoff · API · 1017 v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark174–3Finished6
d1-ultra v DeepSeek-V4.1-Flashoff · API · 1017175–5Finished8
DeepSeek-V4.1-Flashoff · API · 1017 v Claude Fable 5.1low171–8Finished6
DeepSeek-V4.1-Flashoff · API · 1017 v GLiDEAPI174–6Finished6
DeepSeek-V4.1-Flashoff · API · 1017 v Grok 4.7low171–6Finished6
Clef v Gemini 3.8 Flashhigh172–4Finished6
Clef v Claude Sonnet 5.5low172–6Finished4
Claude Sonnet 5.5high v Clef175–1Finished4
Clef v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release170–6Finished6
GLM-5.3low v Clef172–5Finished4
Clef v GNUbg weak172–5Finished6
Qwen3.8 Flashoff v Clef174–5Finished6
Clef v Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark174–6Finished6
Clef v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark170–5Finished4
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Clef175–1Finished4
Clef v d1-ultra171–5Finished4
Claude Fable 5.1low v Clef177–2Finished6
GLiDEAPI v Clef176–1Finished6
Grok 4.7low v Clef177–1Finished4
Clef v DeepSeek-V4.1-Flashoff · API · 1017174–6Finished6
Clef v DeepSeek-V4.1-Flashlow · API · 1017170–8Finished6
Clef v Jev 1.14173–4Finished6
DeepSeek-V4.1-Flashlow · API · 1017 v Gemini 3.8 Flashhigh185–4Finished8
Claude Sonnet 5.5low v DeepSeek-V4.1-Flashlow · API · 1017187–4Finished8
DeepSeek-V4.1-Flashlow · API · 1017 v Claude Sonnet 5.5high182–5Finished6
DeepSeek-V4.1-Flashlow · API · 1017 v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release186–2Finished6
DeepSeek-V4.1-Flashlow · API · 1017 v GLM-5.3low185–2Finished6
DeepSeek-V4.1-Flashlow · API · 1017 v GNUbg weak187–2Finished6
Qwen3.8 Flashoff v DeepSeek-V4.1-Flashlow · API · 1017182–5Finished4
DeepSeek-V4.1-Flashlow · API · 1017 v Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark187–1Finished6
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v DeepSeek-V4.1-Flashlow · API · 1017184–2Finished6
DeepSeek-V4.1-Flashlow · API · 1017 v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark182–5Finished6
DeepSeek-V4.1-Flashlow · API · 1017 v d1-ultra185–3Finished6
Claude Fable 5.1low v DeepSeek-V4.1-Flashlow · API · 1017181–5Finished4
GLiDEAPI v DeepSeek-V4.1-Flashlow · API · 1017182–7Finished6
DeepSeek-V4.1-Flashlow · API · 1017 v Grok 4.7low184–3Finished6
DeepSeek-V4.1-Flashoff · API · 1017 v DeepSeek-V4.1-Flashlow · API · 1017181–7Finished4
Jev 1.14 v Gemini 3.8 Flashhigh181–5Finished4
Claude Sonnet 5.5low v Jev 1.14187–0Finished6
Jev 1.14 v Claude Sonnet 5.5high180–6Finished4
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release v Jev 1.14187–2Finished6
GLM-5.3low v Jev 1.14182–5Finished6
Qwen3.8 Flashoff v Jev 1.14181–5Finished6
Jev 1.14 v Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark183–4Finished6
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Jev 1.14185–1Finished4
Jev 1.14 v Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark181–5Finished6
Jev 1.14 v Claude Fable 5.1low180–6Finished4
Grok 4.7low v Jev 1.14185–2Finished4
DeepSeek-V4.1-Flashoff · API · 1017 v Jev 1.14180–6Finished4
DeepSeek-V4.1-Flashlow · API · 1017 v Jev 1.14183–5Finished6
Gemini 3.8 Flashlow v Gemini 3.8 Flashhigh182–4Finished6
Claude Sonnet 5.5low v Gemini 3.8 Flashlow184–4Finished8
Claude Sonnet 5.5high v Gemini 3.8 Flashlow182–5Finished6
Gemini 3.8 Flashlow v DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release186–2Finished6
GLM-5.3low v Gemini 3.8 Flashlow181–5Finished4
Gemini 3.8 Flashlow v GNUbg weak188–0Finished4
Qwen3.8 Flashoff v Gemini 3.8 Flashlow181–6Finished4
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Gemini 3.8 Flashlow185–1Finished4
Gemini 3.8 Flashlow v Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark187–3Finished6
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark v Gemini 3.8 Flashlow183–5Finished6
Gemini 3.8 Flashlow v d1-ultra185–2Finished6
Claude Fable 5.1low v Gemini 3.8 Flashlow181–5Finished6
GLiDEAPI v Gemini 3.8 Flashlow183–5Finished6
Gemini 3.8 Flashlow v Grok 4.7low185–2Finished6
Gemini 3.8 Flashlow v DeepSeek-V4.1-Flashoff · API · 1017185–3Finished6
Gemini 3.8 Flashlow v DeepSeek-V4.1-Flashlow · API · 1017185–5Finished8
Gemini 3.8 Flashlow v Jev 1.14185–0Finished4
Clef v Gemini 3.8 Flashlow182–5Finished6

This league table's matches. Every match of every stage (ladder rounds, gap matches, late entries, the Hype Cup, the Nackgammon knockout) and the voided ones: Matches.

League 3

FinishedPublished

Swipe for more columns →

Standings: 3 points a win, 1 a draw; only finished league-stage matches count
Model and settingPlacePlayedWon-drawn-lostPointsGames won-lostGammons wonTo playStarting groupLeague moves
GPT-6.1 Sollow1119-0-22741-25110Workhorse original League 2↑ now in League 3round 1: League 2 → 3 (promoted)
GPT-6 Astramediumchat model, one letter · subscription2118-0-32436-26120Frontier original League 3
GPT-6.1 Solhigh3117-0-42138-2680Workhorse original League 2↑ now in League 3round 2: League 2 → 3 (promoted)
Grok 4.7highchat model, one letter · subscription4116-1-41934-3430Frontier original League 3
Claude Opus 5.5highchat model, one letter · subscription5116-0-51839-3170Frontier original League 3
Claude Fable 5.1highchat model, one letter · subscription6116-0-51832-28100Frontier original League 3
GPT-6 Astralowchat model, one letter · subscription7114-1-61330-3070Frontier original League 3
Gemini 4 Argonhighchat model, one letter · subscription8114-1-61327-3950Frontier original League 3
Claude Opus 5.5lowchat model, one letter · subscription9114-0-71231-3770Frontier original League 3
Kimi K3.1 Prooffchat model, one letter · API10114-0-71227-3180Frontier original League 3
Gemini 4 Argonlowchat model, one letter · subscription11113-1-71027-4350Frontier original League 3↕ back in League 3round 3: League 3 → 2 (relegated) · round 4: League 2 → 3 (promoted) · round 5: League 3 → 2 (relegated) · round 6: League 2 → 3 (final swap (ladder ended))
GNUbg weakprogram · CPU12112-2-7826-3870Decision, Workhorse, Frontier program, original Leagues 1, 2, 3

66 of 66 matches finished, 388 games; started 2026-10-20, finished 2026-10-26. 16,355 of 16,355 decisions engine-scored so far.

Notes

  • Won: GPT-6.1 Sollow, 27 points from 9-0-2 in 11 matches.
  • Lowest error rate in league play: GPT-6.1 Solhigh, 9.9 (8.6 to 11.3); no clear difference in league play from GPT-6 Astramedium, GPT-6 Astralow, GPT-6.1 Sollow, Gemini 4 Argonhigh.
  • Claude Opus 5.5, low → high: no clear difference in league play.
  • GPT-6 Astra, low → medium: no clear difference in league play.
  • GPT-6.1 Sol, low → high: no clear difference in league play.
  • Gemini 4 Argon, low → high: no clear difference in league play.
  • League comparisons are not paired: each setup met different positions. Paired comparisons come from the shared positions.
  • The anchor is GNU Backgammon 1.08 Beginner, x86-64 build; runner changes during the leagues are listed in Method.

Decision quality

Swipe for more columns →

Decision quality in this league's games; intervals resample whole games
Model and settingError rate (PR), 95% intervalBand on the human scaleEngine's best playBlunder rateMedian / P90 s per decision$ per 1,000 decisionsDecisions scoredInvalid answers
GPT-6.1 Solhigh9.98.6 to 11.3Advanced (approx.)64.4%7.8%13.1 / 56.515.573API-equivalent (subscription)1,389†242 split into groups0
GPT-6 Astramediumchat model, one letter · subscription11.19.5 to 12.9Advanced (approx.)65.2%8.6%4.40 / 15.150.268API-equivalent (subscription)1,362†182 split into groups0
GPT-6 Astralowchat model, one letter · subscription11.89.5 to 14.2Advanced (approx.)66.3%9.4%3.94 / 14.049.737API-equivalent (subscription)1,251†169 split into groups0
GPT-6.1 Sollow12.110.5 to 13.7Advanced (approx.)65.5%9.4%4.33 / 14.310.228API-equivalent (subscription)1,344†212 split into groups0
Gemini 4 Argonhighchat model, one letter · subscription12.610.5 to 14.8Intermediate (approx.)64.6%8.9%27.5 / 148.664.401API-equivalent (subscription)1,332†208 split into groups36
Claude Fable 5.1highchat model, one letter · subscription14.011.5 to 16.3Intermediate (approx.)64.2%11.7%12.7 / 47.1136.268API-equivalent (subscription)1,286†213 split into groups0
Claude Opus 5.5highchat model, one letter · subscription14.412.1 to 17.0Intermediate (approx.)65.9%11.4%6.35 / 18.536.400API-equivalent (subscription)1,477†178 split into groups0
Claude Opus 5.5lowchat model, one letter · subscription15.413.1 to 18.0Intermediate (approx.)61.1%13.7%3.92 / 10.629.062API-equivalent (subscription)1,411†178 split into groups0
Gemini 4 Argonlowchat model, one letter · subscription16.513.7 to 19.5Intermediate (approx.)60.3%12.4%3.82 / 12.210.204API-equivalent (subscription)1,484†204 split into groups0
Grok 4.7highchat model, one letter · subscription18.815.9 to 21.8Casual (approx.)62.5%14.0%122.3 / 588.6117.379API-equivalent (subscription)1,337†195 split into groups0
Kimi K3.1 Prooffchat model, one letter · API28.024.5 to 31.8Beginner (approx.)51.2%21.7%2.11 / 7.233.982API price1,167†131 split into groups0
GNUbg weakprogram · CPU28.125.0 to 31.4Beginner (approx.)44.5%23.3%0.01 / 0.01Not measured1,4790

† Setups got different inputs: when a position had more legal plays than a setup takes in one request, its plays were split into groups and the group winners compared. The rows marked † had such decisions; the others saw every play at once. Every setup got the same long rules text.

Time and cost

Swipe for more columns →

Time and cost in this league
Model and setting$ per 1,000 decisionsMedian / P90 s per decisionMean retry/wait sOver the time budgetTokens in / out per decisionCost coverageDecisions
GPT-6.1 Solhigh15.573API-equivalent (subscription)13.1 / 56.55.6924.5%5,217 / 514100%1,389
GPT-6 Astramediumchat model, one letter · subscription50.268API-equivalent (subscription)4.40 / 15.10.412.6%4,471 / 111100%1,362
GPT-6 Astralowchat model, one letter · subscription49.737API-equivalent (subscription)3.94 / 14.00.231.6%4,486 / 98100%1,251
GPT-6.1 Sollow10.228API-equivalent (subscription)4.33 / 14.30.441.5%4,543 / 114100%1,344
Gemini 4 Argonhighchat model, one letter · subscription64.401API-equivalent (subscription)27.5 / 148.66.9347.1%5,117 / 5,417100%1,368
Claude Fable 5.1highchat model, one letter · subscription136.268API-equivalent (subscription)12.7 / 47.10.5019.4%6,502 / 1,425100%1,286
Claude Opus 5.5highchat model, one letter · subscription36.400API-equivalent (subscription)6.35 / 18.50.514.3%5,861 / 648100%1,477
Claude Opus 5.5lowchat model, one letter · subscription29.062API-equivalent (subscription)3.92 / 10.60.030.6%5,896 / 274100%1,411
Gemini 4 Argonlowchat model, one letter · subscription10.204API-equivalent (subscription)3.82 / 12.20.121.3%4,970 / 27100%1,484
Grok 4.7highchat model, one letter · subscription117.379API-equivalent (subscription)122.3 / 588.62.2185.9%6,400 / 17,430100%1,337
Kimi K3.1 Prooffchat model, one letter · API3.982API price2.11 / 7.230.030.1%4,353 / 20100%1,167
GNUbg weakprogram · CPUNot measured0.01 / 0.010.000.0%Not measured0%1,479
Matches (66)

Swipe for more columns →

Matches
MatchMatchdayScoreStatusGames
Claude Fable 5.1high v Claude Opus 5.5high16–5Finished6
Gemini 4 Argonhigh v Grok 4.7high14–3Finished6
GPT-6 Astralow v GPT-6 Astramedium22–5Finished6
Claude Fable 5.1high v Grok 4.7high25–1Finished6
GPT-6 Astramedium v Claude Opus 5.5low32–5Finished6
GNUbg weak v GPT-6 Astralow31–5Finished4
Gemini 4 Argonhigh v GPT-6 Astralow45–2Finished6
GNUbg weak v GPT-6 Astramedium44–5Finished6
Claude Opus 5.5high v Claude Opus 5.5low45–2Finished6
Claude Opus 5.5low v Grok 4.7high53–4Finished6
Gemini 4 Argonhigh v GNUbg weak54–4Finished8
Claude Opus 5.5high v GPT-6 Astramedium54–2Finished6
GPT-6 Astralow v Claude Fable 5.1high55–2Finished6
Grok 4.7high v Claude Opus 5.5high64–2Finished6
GNUbg weak v Claude Fable 5.1high60–6Finished4
Gemini 4 Argonhigh v GPT-6 Astramedium65–6Finished8
Grok 4.7high v GPT-6 Astramedium71–5Finished4
Claude Fable 5.1high v Gemini 4 Argonhigh75–0Finished4
GPT-6 Astramedium v Claude Fable 5.1high85–3Finished4
GPT-6 Astralow v Claude Opus 5.5low85–0Finished4
GPT-6.1 Sollow v GNUbg weak85–4Finished6
GPT-6 Astralow v Claude Opus 5.5high92–5Finished6
GNUbg weak v Claude Opus 5.5low96–5Finished8
GNUbg weak v GPT-6.1 Solhigh95–1Finished4
GPT-6.1 Sollow v GPT-6.1 Solhigh95–2Finished6
Gemini 4 Argonhigh v Claude Opus 5.5low102–4Finished6
GNUbg weak v Claude Opus 5.5high101–6Finished6
Grok 4.7high v GPT-6 Astralow104–2Finished6
GPT-6 Astralow v GPT-6.1 Sollow101–5Finished4
GPT-6 Astramedium v GPT-6.1 Sollow105–1Finished6
GPT-6.1 Sollow v Claude Opus 5.5low107–6Finished8
Claude Opus 5.5high v GPT-6.1 Sollow105–3Finished8
GPT-6.1 Sollow v Claude Fable 5.1high108–1Finished6
Gemini 4 Argonhigh v GPT-6.1 Sollow102–5Finished6
Grok 4.7high v GPT-6.1 Sollow100–6Finished6
Claude Opus 5.5low v Claude Fable 5.1high113–4Finished6
Gemini 4 Argonhigh v Claude Opus 5.5high115–4Finished6
GNUbg weak v Grok 4.7high115–5Finished8
GPT-6 Astralow v GPT-6.1 Solhigh114–6Finished6
GPT-6 Astramedium v GPT-6.1 Solhigh113–5Finished6
Claude Opus 5.5low v GPT-6.1 Solhigh111–5Finished6
GPT-6.1 Solhigh v Claude Opus 5.5high113–4Finished6
GPT-6.1 Solhigh v Claude Fable 5.1high114–2Finished6
GPT-6.1 Solhigh v Gemini 4 Argonhigh116–0Finished4
GPT-6.1 Solhigh v Grok 4.7high113–5Finished8
Gemini 4 Argonlow v GPT-6 Astralow115–5Finished8
Gemini 4 Argonlow v GPT-6 Astramedium110–5Finished4
Claude Opus 5.5low v Gemini 4 Argonlow114–2Finished6
Claude Opus 5.5high v Gemini 4 Argonlow113–5Finished8
Claude Fable 5.1high v Gemini 4 Argonlow115–6Finished8
Gemini 4 Argonlow v Gemini 4 Argonhigh112–4Finished6
Gemini 4 Argonlow v Grok 4.7high111–5Finished6
GNUbg weak v Gemini 4 Argonlow113–4Finished6
GPT-6.1 Sollow v Gemini 4 Argonlow114–3Finished6
Gemini 4 Argonlow v GPT-6.1 Solhigh112–5Finished6
Kimi K3.1 Prooff v Gemini 4 Argonlow115–2Finished6
Kimi K3.1 Prooff v GPT-6 Astralow122–5Finished4
GPT-6 Astramedium v Kimi K3.1 Prooff126–2Finished6
Kimi K3.1 Prooff v Claude Opus 5.5low124–5Finished6
Kimi K3.1 Prooff v Claude Opus 5.5high124–3Finished6
Kimi K3.1 Prooff v Claude Fable 5.1high121–5Finished4
Kimi K3.1 Prooff v Gemini 4 Argonhigh127–1Finished6
Grok 4.7high v Kimi K3.1 Prooff125–2Finished6
GNUbg weak v Kimi K3.1 Prooff120–6Finished4
GPT-6.1 Sollow v Kimi K3.1 Prooff125–1Finished4
GPT-6.1 Solhigh v Kimi K3.1 Prooff128–1Finished6

This league table's matches. Every match of every stage (ladder rounds, gap matches, late entries, the Hype Cup, the Nackgammon knockout) and the voided ones: Matches.

The leagues as they stand now, after 24 moves in ladder rounds 1–6 (the ladder ended on 26 Oct; its final swap stands). These leagues mix starting groups, so they keep their numbers: League 1, League 2, League 3. Each row says its starting group and, for a setup that moved, its recorded moves. The leaderboard's League keys and links such as ?league=league-3 refer to these leagues.

Cups

Hype Cup

Finished

Quarter-finals

  1. Quarter-final 1Played

    1. 1GPT-6 Astramedium (winner)6
    2. 8Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1

    2026-10-29: match #770

  2. Quarter-final 2Played

    1. 2Claude Fable 5.1high (winner)6
    2. 6GLiDEAPI0

    2026-10-28: match #767

  3. Quarter-final 3Stored

    1. 3GPT-6.1 Sollow (winner)6
    2. 7Grok 4.7high0

    Played in Ladder round 1, League 3, 2026-10-24: match #550

  4. Quarter-final 4Stored

    1. 4GPT-6.1 Solhigh (winner)7
    2. 5d1-ultra0

    Played in Ladder round 1, League 2, 2026-10-23: match #512

Semi-finals

  1. Semi-final AStored

    1. 1GPT-6 Astramedium3
    2. 4GPT-6.1 Solhigh (winner)5

    Played in Ladder round 2, League 3, 2026-10-24: match #555

  2. Semi-final BStored

    1. 2Claude Fable 5.1high1
    2. 3GPT-6.1 Sollow (winner)8

    Played in Ladder round 1, League 3, 2026-10-23: match #516

Final

  1. FinalStored

    1. 4GPT-6.1 Solhigh2
    2. 3GPT-6.1 Sollow (winner)5

    Played in League 2 season, 2026-10-19: match #126

Played (2): a new Cup match. Stored (5): Two setups play each other once: a pairing already played in a league stands, and the bracket shows where it was played. A level match goes to the setup with the lower error rate in that match.

Roster. Teams from the original-league tables: each setup in its original league, only matches within it count (3 points a win, 1 a draw, 0 a loss; tiebreak: head-to-head, then the lower error rate; rule set on 28 Oct, after the ties were known). The top 2 models of each table and the two best 3rd places by error rate; programs and late entries are not eligible.

Nackgammon knockout

Finished

Quarter-finals

  1. Quarter-final 1Played

    1. 1GPT-6 Astramedium (winner)4
    2. 8Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark2

    2026-10-29: knockout match #3

  2. Quarter-final 2Played

    1. 2Claude Fable 5.1high (winner)6
    2. 6GLiDEAPI0

    2026-10-28: knockout match #1

  3. Quarter-final 3Played

    1. 3GPT-6.1 Sollow5
    2. 7Grok 4.7high (winner)8

    2026-10-29: knockout match #4

  4. Quarter-final 4Played

    1. 4GPT-6.1 Solhigh (winner)6
    2. 5d1-ultra1

    2026-10-28: knockout match #2

Semi-finals

  1. Semi-final APlayed

    1. 1GPT-6 Astramedium2
    2. 4GPT-6.1 Solhigh (winner)5

    2026-10-29: knockout match #5

  2. Semi-final BPlayed

    1. 2Claude Fable 5.1high (winner)5
    2. 7Grok 4.7high4

    2026-10-29: knockout match #6

Final

  1. FinalPlayed

    1. 4GPT-6.1 Solhigh (winner)5
    2. 2Claude Fable 5.1high4

    2026-10-29: knockout match #7

Nackgammon start; all 7 matches new (stored standard results don't count). The same 8 teams and the same draw, every game from the Nackgammon start; otherwise the league match rules.

Every match of both, with the leagues' matches: Matches. The Cup is an event. The error rate is the ranking.

Knockout events, apart from the league tables: their places and results are not league points. The Hype Cup took its 8 teams from the starting groups' tables (its roster rule is under its bracket); the Nackgammon knockout used the same 8 teams and the same draw. A Cup place is not a promotion.

Late entries (not in standings)

Swipe for more columns →

Late entries: each one's own late-entry games; never ranked
SetupEntry dateLeague field playedMatches playedRecord (W–D–L)Error rate (PR), 95% intervalRisk judgement, margin 0.001Ability d′Caution c (+ cautious, − bold)Engine's best playMedian s per decision$ per 1,000 decisionsInvalid answersCut off at the output limitRoute and effort sent
GPT-6 AstrahighPublished2026-10-28first match 15:35 UTCLeague 3 (original table)12 setups12 / 1210–1–1won–drawn–lost8.87.4 to 10.374.471.8 to 77.01.871.68 to 2.08 · risks taken 267 / 465 Risk pays · needless 29 / 643 Safety pays+0.75+0.64 to +0.87 · cautious67.6%9.8359.356API-equivalent (subscription), $10.0 per million input tokens, $50.0 per million output tokens0of 1,489 decisions00 of 1,489 decisions sent an output limit; 0 over 32,768 output tokens without onechatgpt-subscriptionopenai/gpt-6-astra · effort high in 1,489 of 1,489 decisions
SubQ Snap 1.1PublishedThis pair may be SubQ Snap 1.1 against itself: under the 'subq-snap-preview' name the API returned answers identical to SubQ Snap 1.1 (1,000 identical paired positions: letters and probabilities).2026-11-02first match 18:29 UTCLeague 1 (original table)19 setups19 / 1912–0–7won–drawn–lost40.236.7 to 43.856.754.8 to 58.60.710.61 to 0.82 · risks taken 412 / 798 Risk pays · needless 253 / 1,003 Safety pays+0.31+0.25 to +0.38 · cautious42.9%0.230.150API price (billed); output free, $0.04 per million input tokens; docs.subq.ai/using-the-api/models, read 2026-11-010of 2,326 decisions00 of 2,326 decisions sent an output limit; 0 over 32,768 output tokens without onesubqsubq/subq-snap-1.1 · effort not in the requests
Elephant-1Published2026-11-02first match 18:29 UTCLeague 1 (original table)19 setups19 / 1911–0–8won–drawn–lost44.640.3 to 48.957.755.8 to 59.60.910.80 to 1.04 · risks taken 337 / 812 Risk pays · needless 102 / 784 Safety pays+0.67+0.60 to +0.74 · cautious41.8%0.260.685estimate · RTX PRO 6000 Blackwell0of 2,044 decisions00 of 2,044 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (baselabs_server.py)baselabs/baselabs/elephant-1 · effort not in the requests
Mammoth-1Published2026-11-02first match 18:29 UTCLeague 1 (original table)19 setups19 / 1916–1–2won–drawn–lost28.526.2 to 31.163.261.2 to 65.21.100.97 to 1.23 · risks taken 361 / 720 Risk pays · needless 120 / 880 Safety pays+0.55+0.48 to +0.62 · cautious49.0%0.720.999estimate · RTX PRO 6000 Blackwell0of 2,021 decisions00 of 2,021 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (baselabs_server.py)baselabs/baselabs/mammoth-1 · effort not in the requests
Drex Max v1.1Published2026-11-06first match 00:49 UTCLeague 1 (original table)19 setups18 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)12–2–4won–drawn–lost30.928.2 to 33.759.157.0 to 61.30.740.60 to 0.87 · risks taken 297 / 702 Risk pays · needless 164 / 939 Safety pays+0.56+0.49 to +0.64 · cautious46.8%0.440.265list price; output not billed, $0.06 per million input tokens; nace.ai/drex and console.nace.ai/docs/drex/reference/pricing, read 2026-11-05billed at Nace's launch price, $0.03 per million input tokens: 0.1320of 2,154 decisions00 of 2,154 decisions sent an output limit; 0 over 32,768 output tokens without onenacenace/drex-max-v1.1 · effort not in the requests
Drex 2.0Published2026-11-06first match 00:49 UTCLeague 1 (original table)19 setups18 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)10–0–8won–drawn–lost38.435.6 to 41.458.155.7 to 60.40.890.73 to 1.04 · risks taken 311 / 699 Risk pays · needless 112 / 735 Safety pays+0.58+0.51 to +0.65 · cautious45.2%0.420.258list price (billed); output not billed, $0.06 per million input tokens; console.nace.ai/docs/drex/reference/pricing, read 2026-11-050of 1,937 decisions00 of 1,937 decisions sent an output limit; 0 over 32,768 output tokens without onenacenace/drex-v2.0 · effort not in the requests
Drex MaxNVIDIA RTX PRO 6000 BlackwellPublished2026-11-06first match 01:02 UTCLeague 1 (original table)19 setups18 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)11–0–7won–drawn–lost30.427.4 to 33.462.961.1 to 64.71.080.95 to 1.22 · risks taken 373 / 761 Risk pays · needless 121 / 896 Safety pays+0.56+0.48 to +0.65 · cautious48.2%0.681.285estimate · RTX PRO 6000 Blackwell0of 2,148 decisions00 of 2,148 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (drex_max_server.py)drex-max/drex-max-v1 · effort not in the requests
Drex 2.0NVIDIA RTX PRO 6000 BlackwellPublished2026-11-06first match 07:44 UTCLeague 1 (original table)19 setups18 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)7–0–11won–drawn–lost42.939.1 to 47.153.050.5 to 55.20.570.43 to 0.71 · risks taken 286 / 720 Risk pays · needless 148 / 729 Safety pays+0.54+0.47 to +0.63 · cautious42.7%0.400.782estimate · RTX PRO 6000 Blackwell0of 1,993 decisions00 of 1,993 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (drex-v2.0 serve.py)drex-local/drex-v2.0 · effort not in the requests
REAX 1.3Published2026-11-07first match 11:33 UTCLeague 1 (original table)19 setups18 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)11–1–6won–drawn–lost27.824.9 to 30.660.558.7 to 62.40.930.78 to 1.06 · risks taken 249 / 666 Risk pays · needless 90 / 850 Safety pays+0.78+0.70 to +0.87 · cautious51.7%1.741.545estimate · RTX PRO 6000 Blackwell0of 1,998 decisions00 of 1,998 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (reax13 serve.py)reax/reax13 · effort not in the requests
Muse EmbermaxPublished2026-11-10first match 14:44 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)13–0–4won–drawn–lost20.117.6 to 22.766.063.9 to 68.21.341.17 to 1.53 · risks taken 241 / 596 Risk pays · needless 40 / 711 Safety pays+0.91+0.83 to +1.01 · cautious58.1%15.36.863API price (billed), $0.1 per million input tokens, $0.5 per million output tokens; dev.meta.ai/models/muse-ember and OpenRouter's endpoints list for meta/muse-ember (Meta: $0.10/M prompt, $0.50/M completion), read 2026-11-10, read 2026-11-1046of 1,713 decisions451,713 of 1,713 decisions sent an output limit; 0 over 32,768 output tokens without oneopenrouter:metaopenai-compatible/meta/muse-ember · effort max in 1,713 of 1,713 decisions
KnowLine-31B-Gen1.5NVIDIA DGX Spark (GB10)Published2026-11-10first match 20:35 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)13–1–3won–drawn–lost26.223.0 to 29.461.659.6 to 63.51.140.96 to 1.33 · risks taken 212 / 649 Risk pays · needless 45 / 803 Safety pays+1.02+0.94 to +1.11 · cautious54.6%3.45No priceno price (our GPU); no list price published for KnowLine-31B-Gen1.5, read 2026-11-100of 1,928 decisions00 of 1,928 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (knowline_server.py)knowline/m · effort not in the requests
KnowLine 4B Gen5.5Published2026-11-11first match 11:48 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)5–1–11won–drawn–lost53.648.8 to 58.649.747.0 to 52.30.590.47 to 0.72 · risks taken 448 / 778 Risk pays · needless 281 / 813 Safety pays+0.10+0.02 to +0.18 · cautious36.9%0.280.054estimate · DGX Spark0of 2,153 decisions00 of 2,153 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (knowline_server.py)knowline/m · effort not in the requests
StartLux-Decision-v1.1-2B-InstantPublished2026-11-11first match 12:42 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)3–0–14won–drawn–lost50.045.4 to 54.853.851.3 to 56.60.700.54 to 0.86 · risks taken 330 / 630 Risk pays · needless 158 / 608 Safety pays+0.29+0.20 to +0.38 · cautious36.2%0.210.052estimate · DGX Spark0of 1,771 decisions00 of 1,771 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (startlux-decide serve)startlux · effort not in the requests
KnowLine-0.8B-Gen1Published2026-11-11first match 18:27 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)1–1–15won–drawn–lost71.966.2 to 77.945.242.5 to 47.70.480.33 to 0.62 · risks taken 379 / 637 Risk pays · needless 240 / 594 Safety pays+0.00−0.07 to +0.08 · cautious30.3%0.110.049estimate · DGX Spark0of 1,728 decisions00 of 1,728 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (knowline_server.py)knowline/m · effort not in the requests
Gemini Decision 1Published2026-11-11first match 19:28 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)13–0–4won–drawn–lost28.025.0 to 31.164.762.2 to 67.21.271.11 to 1.46 · risks taken 283 / 535 Risk pays · needless 84 / 732 Safety pays+0.56+0.48 to +0.66 · cautious52.3%0.240.382API price (billed), $0.1 per million input tokens; aistudio.google.com/docs/pricing#gemini-decision and openrouter.ai/google/gemini-decision-1 ($0.10/M input, no output tokens), read 2026-11-11; cost is the billed cost per decision (turns.cost_usd), read 2026-11-110of 1,704 decisions00 of 1,704 decisions sent an output limit; 0 over 32,768 output tokens without oneopenrouter (google ai studio)openrouter-decisions/google/gemini-decision-1 · effort not in the requests
StartLux-Decision-v1.1-0.8B-InstantPublished2026-11-11first match 19:49 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)3–0–14won–drawn–lost64.759.3 to 70.344.642.3 to 46.90.650.50 to 0.80 · risks taken 488 / 665 Risk pays · needless 351 / 715 Safety pays−0.30−0.36 to −0.24 · bold32.5%0.130.054estimate · DGX Spark0of 1,882 decisions00 of 1,882 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (startlux-decide serve)startlux · effort not in the requests
OneJev1.6-0.8BPublished2026-11-11first match 21:19 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)0–0–17won–drawn–lost82.575.2 to 90.842.539.3 to 45.80.450.27 to 0.61 · risks taken 365 / 606 Risk pays · needless 270 / 634 Safety pays−0.04−0.11 to +0.05 · bold28.1%0.100.052estimate · DGX Spark0of 1,763 decisions00 of 1,763 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (flide serve)onejev · effort not in the requests
GLiClass v4 EdgePublished2026-11-12first match 08:43 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)0–0–17won–drawn–lost133.5118.5 to 148.821.518.6 to 24.60.340.13 to 0.57 · risks taken 461 / 508 Risk pays · needless 400 / 478 Safety pays−1.15−1.29 to −1.02 · bold15.7%1.330.174estimate · DGX Spark0of 1,405 decisions00 of 1,405 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (gliclass_server.py)gliclass · effort not in the requests
KnowLine-2B-Gen1Published2026-11-12first match 16:50 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)3–0–14won–drawn–lost54.548.9 to 60.851.148.7 to 53.50.590.45 to 0.72 · risks taken 337 / 676 Risk pays · needless 195 / 703 Safety pays+0.30+0.22 to +0.38 · cautious35.6%0.150.049estimate · DGX Spark0of 1,955 decisions00 of 1,955 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (knowline_server.py)knowline/m · effort not in the requests
GLiDE 2BPublished2026-11-12first match 16:58 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)1–0–16won–drawn–lost67.461.0 to 74.546.743.6 to 49.60.650.49 to 0.81 · risks taken 378 / 581 Risk pays · needless 233 / 587 Safety pays−0.06−0.15 to +0.03 · bold30.3%0.330.058estimate · DGX Spark0of 1,676 decisions00 of 1,676 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (glide.server)glide · effort not in the requests
GLiDE 0.8BPublished2026-11-12first match 17:08 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)2–1–14won–drawn–lost80.471.8 to 89.341.738.9 to 44.30.580.44 to 0.72 · risks taken 464 / 671 Risk pays · needless 311 / 663 Safety pays−0.21−0.29 to −0.13 · bold26.6%0.220.051estimate · DGX Spark0of 1,832 decisions00 of 1,832 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (glide.server)glide · effort not in the requests
GLiDE 0.4BPublished2026-11-12first match 17:18 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)1–0–16won–drawn–lost113.2102.1 to 124.128.925.6 to 32.10.630.42 to 0.85 · risks taken 361 / 441 Risk pays · needless 304 / 498 Safety pays−0.59−0.72 to −0.49 · bold17.7%0.130.057estimate · DGX Spark0of 1,355 decisions00 of 1,355 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (glide.server)glide-0.4b · effort not in the requests
Sieve 31B v3Published2026-11-12first match 19:29 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)15–0–2won–drawn–lost27.524.1 to 31.662.760.6 to 64.91.040.88 to 1.22 · risks taken 228 / 568 Risk pays · needless 76 / 772 Safety pays+0.77+0.69 to +0.85 · cautious53.8%11.10.794estimate · DGX Spark0of 1,740 decisions00 of 1,740 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (sieve-serve)sieve · effort not in the requests
Turion Jev 31BPublished2026-11-13first match 12:30 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)13–1–3won–drawn–lost29.726.5 to 33.359.657.7 to 61.41.050.87 to 1.26 · risks taken 184 / 674 Risk pays · needless 35 / 715 Safety pays+1.13+1.04 to +1.24 · cautious51.8%3.950.293estimate · DGX Spark0of 1,855 decisions00 of 1,855 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (turion_jev.serve)turion-jev · effort not in the requests
StartLux-Decision-v1.1-4B-InstantPublished2026-11-13first match 16:16 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)10–0–7won–drawn–lost40.136.7 to 43.757.255.0 to 59.30.790.67 to 0.92 · risks taken 370 / 777 Risk pays · needless 145 / 739 Safety pays+0.46+0.36 to +0.55 · cautious42.4%0.440.072estimate · DGX Spark0of 1,989 decisions00 of 1,989 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (startlux-decide serve)startlux · effort not in the requests
Cerno 9B v1.2Published2026-11-13first match 17:22 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)6–1–10won–drawn–lost48.544.5 to 52.656.153.9 to 58.20.820.69 to 0.97 · risks taken 428 / 734 Risk pays · needless 232 / 857 Safety pays+0.20+0.13 to +0.27 · cautious40.6%1.960.139estimate · DGX Spark0of 2,108 decisions00 of 2,108 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (cerno_server.py around CernoEngine)cerno · effort not in the requests
d1-2BNVIDIA RTX PRO 6000 BlackwellPublished2026-11-14first match 16:12 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)5–0–12won–drawn–lost47.142.8 to 51.555.252.8 to 57.50.750.62 to 0.89 · risks taken 349 / 688 Risk pays · needless 180 / 776 Safety pays+0.36+0.29 to +0.43 · cautious39.7%0.060.467estimate · RTX PRO 6000 Blackwell0of 1,980 decisions00 of 1,980 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (d1_server.py)d1-local · effort not in the requests
d1-Omni-2BNVIDIA RTX PRO 6000 BlackwellPublished2026-11-14first match 19:47 UTCLeague 1 (original table)19 setups17 / 19GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)Mercury Decide: outside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)9–0–8won–drawn–lost49.843.5 to 57.453.250.9 to 55.60.610.48 to 0.75 · risks taken 287 / 619 Risk pays · needless 175 / 728 Safety pays+0.40+0.32 to +0.48 · cautious40.4%0.060.657estimate · RTX PRO 6000 Blackwell0of 1,710 decisions00 of 1,710 decisions sent an output limit; 0 over 32,768 output tokens without onelocal (d1_server.py)d1-local/LiquidAI/d1-Omni-2B · effort not in the requests

A late entry joins an original league table after it was played and plays every setup of that table once, in matches of its own. Its games count in no standings table, no league table, no Hype Cup input and no other setup's numbers. The numbers here come from its own late-entry games only, computed as for a setup in a league. Late entries are ranked on the risk leaderboard; they do not change league match-points standings, Cups, or other setups' results.

GPT-6 Astrahigh: results per opponent

Swipe for more columns →

GPT-6 Astrahigh against each opponent, its score first: 10 won, 1 drawn, 1 lost; none counts in a standings table
OpponentScoreResult#Finished
GNUbg weak5–0Won7582026-10-28
GPT-6 Astralow5–2Won7592026-10-28
GPT-6 Astramedium4–3Won7602026-10-28
Gemini 4 Argonhigh5–1Won7612026-10-28
Gemini 4 Argonlow6–3Won7622026-10-28
Claude Fable 5.1low5–1Won7632026-10-28
Claude Fable 5.1high0–7Lost7642026-10-28
Claude Opus 5.5low4–3Won7652026-10-28
Claude Opus 5.5high5–5Draw7662026-10-28
Grok 4.7high4–2Won7682026-10-29
Grok 4.7low6–1Won7692026-10-28
Kimi K3.1 Prooff6–2Won7712026-10-30

SubQ Snap 1.1: results per opponent

Swipe for more columns →

SubQ Snap 1.1 against each opponent, its score first: 12 won, 0 drawn, 7 lost; none counts in a standings table
OpponentScoreResult#Finished
SubQ Snap PreviewThis pair may be SubQ Snap 1.1 against itself: under the 'subq-snap-preview' name the API returned answers identical to SubQ Snap 1.1 (1,000 identical paired positions: letters and probabilities).7–4Won7722026-11-02
Jev 1.144–2Won7732026-11-02
celeris-1-decision0–6Lost7742026-11-02
GLiDEAPI2–5Lost7752026-11-02
Jev 1.136–0Won7842026-11-02
GPT-6 LunaDecisions API4–5Lost7852026-11-02
Safest6–2Won7862026-11-02
Mercury Decide5–1Won7872026-11-02
Winnow-12B5–0Won7882026-11-02
d1-ultra2–5Lost7892026-11-02
GNUbg weak0–7Lost7902026-11-02
Random9–0Won7912026-11-02
Perplexity Decider Flash v1.25–3Won7922026-11-02
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell4–2Won7932026-11-02
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell2–7Lost7952026-11-02
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–3Won8172026-11-02
Clef2–6Lost8222026-11-02
Cygnet5–2Won8242026-11-02
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell7–1Won8262026-11-02

Elephant-1: results per opponent

Swipe for more columns →

Elephant-1 against each opponent, its score first: 11 won, 0 drawn, 8 lost; none counts in a standings table
OpponentScoreResult#Finished
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.2–6Lost7762026-11-02
Jev 1.143–5Lost7772026-11-02
celeris-1-decision4–3Won7782026-11-02
GLiDEAPI5–2Won7792026-11-02
Jev 1.132–4Lost7942026-11-02
GPT-6 LunaDecisions API5–0Won7962026-11-02
Safest5–3Won7972026-11-02
Mercury Decide6–1Won7982026-11-02
Winnow-12B1–5Lost7992026-11-02
d1-ultra3–4Lost8002026-11-02
GNUbg weak1–5Lost8012026-11-02
Random8–0Won8022026-11-02
Perplexity Decider Flash v1.25–3Won8032026-11-02
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell5–2Won8042026-11-02
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell3–4Lost8052026-11-02
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–3Won8192026-11-02
Cygnet5–6Lost8232026-11-02
Clef6–1Won8252026-11-02
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell7–2Won8272026-11-02

Mammoth-1: results per opponent

Swipe for more columns →

Mammoth-1 against each opponent, its score first: 16 won, 1 drawn, 2 lost; none counts in a standings table
OpponentScoreResult#Finished
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.5–2Won7802026-11-02
celeris-1-decision4–2Won7812026-11-02
GLiDEAPI5–2Won7822026-11-02
Jev 1.144–3Won7832026-11-02
Jev 1.135–2Won8062026-11-02
GPT-6 LunaDecisions API6–4Won8072026-11-02
Safest5–5Draw8082026-11-02
Mercury Decide6–2Won8092026-11-02
Winnow-12B5–1Won8102026-11-02
d1-ultra4–5Lost8112026-11-02
GNUbg weak2–6Lost8122026-11-02
Random5–0Won8132026-11-02
Perplexity Decider Flash v1.25–1Won8142026-11-02
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell4–2Won8152026-11-02
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell4–3Won8162026-11-02
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark5–1Won8182026-11-02
Cygnet6–0Won8202026-11-02
Clef5–1Won8212026-11-02
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–1Won8282026-11-02

Drex Max v1.1: results per opponent

Swipe for more columns →

Drex Max v1.1 against each opponent, its score first: 12 won, 2 drawn, 4 lost; 1 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
Jev 1.135–5Draw8302026-11-06
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.3–5Lost8312026-11-06
Mercury Decide5–0Won8322026-11-06
GLiDEAPI2–5Lost8392026-11-06
Random6–0Won8422026-11-06
celeris-1-decision6–1Won8442026-11-06
Jev 1.147–2Won8502026-11-06
GNUbg weak4–3Won8512026-11-06
Safest6–2Won8542026-11-06
d1-ultra4–3Won8622026-11-06
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell4–5Lost8652026-11-06
Perplexity Decider Flash v1.26–1Won8662026-11-06
Winnow-12B5–2Won8692026-11-06
Cygnet3–4Lost8712026-11-06
Clef6–5Won8752026-11-06
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell6–6Draw8782026-11-06
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–0Won8802026-11-06
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark5–4Won8842026-11-06
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A

Drex 2.0: results per opponent

Swipe for more columns →

Drex 2.0 against each opponent, its score first: 10 won, 0 drawn, 8 lost; 1 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.0–5Lost8332026-11-06
Mercury Decide5–2Won8342026-11-06
Jev 1.136–2Won8362026-11-06
GLiDEAPI3–5Lost8372026-11-06
Random5–0Won8382026-11-06
celeris-1-decision1–5Lost8402026-11-06
Jev 1.145–3Won8412026-11-06
GNUbg weak5–2Won8432026-11-06
Safest3–5Lost8492026-11-06
d1-ultra3–5Lost8522026-11-06
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell5–2Won8532026-11-06
Perplexity Decider Flash v1.25–1Won8582026-11-06
Winnow-12B7–6Won8602026-11-06
Cygnet1–5Lost8632026-11-06
Clef0–5Lost8762026-11-06
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell7–4Won8772026-11-06
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–3Won8812026-11-06
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0–5Lost8832026-11-06
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A

Drex MaxNVIDIA RTX PRO 6000 Blackwell: results per opponent

Swipe for more columns →

Drex MaxNVIDIA RTX PRO 6000 Blackwell against each opponent, its score first: 11 won, 0 drawn, 7 lost; 1 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
Mercury Decide5–1Won8452026-11-06
Jev 1.133–4Lost8472026-11-06
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.5–4Won8482026-11-06
GLiDEAPI3–5Lost8552026-11-06
Random5–0Won8562026-11-06
celeris-1-decision4–2Won8572026-11-06
Jev 1.143–5Lost8592026-11-06
GNUbg weak0–5Lost8612026-11-06
Safest5–3Won8642026-11-06
d1-ultra6–1Won8672026-11-06
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell2–5Lost8682026-11-06
Perplexity Decider Flash v1.26–2Won8702026-11-06
Winnow-12B7–1Won8722026-11-06
Cygnet3–4Lost8732026-11-06
Clef8–3Won8742026-11-06
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell5–3Won8792026-11-06
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell6–0Won8822026-11-06
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark3–4Lost8852026-11-06
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A

Drex 2.0NVIDIA RTX PRO 6000 Blackwell: results per opponent

Swipe for more columns →

Drex 2.0NVIDIA RTX PRO 6000 Blackwell against each opponent, its score first: 7 won, 0 drawn, 11 lost; 1 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.2–5Lost8872026-11-06
Jev 1.133–5Lost8882026-11-06
Mercury Decide6–2Won8892026-11-06
GLiDEAPI4–3Won8902026-11-06
Random8–0Won8912026-11-06
celeris-1-decision3–5Lost8922026-11-06
Jev 1.142–4Lost8932026-11-06
GNUbg weak0–7Lost8942026-11-06
Safest5–3Won8952026-11-06
d1-ultra1–5Lost8962026-11-06
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell0–5Lost8972026-11-06
Perplexity Decider Flash v1.26–3Won8982026-11-06
Winnow-12B2–5Lost8992026-11-06
Cygnet2–6Lost9002026-11-06
Clef2–5Lost9012026-11-06
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell5–2Won9022026-11-06
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–0Won9032026-11-06
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–5Lost9042026-11-06
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A

REAX 1.3: results per opponent

Swipe for more columns →

REAX 1.3 against each opponent, its score first: 11 won, 1 drawn, 6 lost; 1 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
Random7–0Won9052026-11-07
GNUbg weak1–5Lost9062026-11-07
Safest6–0Won9072026-11-07
Cygnet2–5Lost9082026-11-07
Winnow-12B6–3Won9092026-11-07
Perplexity Decider Flash v1.25–0Won9102026-11-07
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell6–0Won9112026-11-07
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell6–2Won9122026-11-07
Clef6–0Won9132026-11-07
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–1Won9142026-11-07
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1–5Lost9152026-11-07
celeris-1-decision4–5Lost9162026-11-07
GLiDEAPI5–5Draw9172026-11-07
Jev 1.135–0Won9182026-11-07
Jev 1.145–6Lost9192026-11-07
d1-ultra5–3Won9202026-11-07
Mercury Decide6–1Won9212026-11-07
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.2–6Lost9222026-11-07
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A

Muse Embermax: results per opponent

Swipe for more columns →

Muse Embermax against each opponent, its score first: 13 won, 0 drawn, 4 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI4–2Won9232026-11-10
celeris-1-decision2–4Lost9242026-11-10
GNUbg weak2–6Lost9252026-11-10
Random5–0Won9262026-11-10
Safest7–1Won9272026-11-10
d1-ultra7–1Won9282026-11-10
Jev 1.145–2Won9292026-11-10
Jev 1.137–2Won9302026-11-10
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.3–4Lost9312026-11-10
Cygnet7–1Won9552026-11-11
Winnow-12B7–0Won9562026-11-11
Perplexity Decider Flash v1.26–0Won9672026-11-11
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell6–0Won9682026-11-11
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell3–4Lost9692026-11-11
Clef4–2Won9802026-11-11
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–1Won9812026-11-11
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark5–2Won9822026-11-11
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

KnowLine-31B-Gen1.5NVIDIA DGX Spark (GB10): results per opponent

Swipe for more columns →

KnowLine-31B-Gen1.5NVIDIA DGX Spark (GB10) against each opponent, its score first: 13 won, 1 drawn, 3 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI4–3Won9372026-11-10
celeris-1-decision2–5Lost9382026-11-10
GNUbg weak1–5Lost9392026-11-10
Random8–0Won9402026-11-10
Safest6–2Won9412026-11-10
d1-ultra3–4Lost9422026-11-10
Jev 1.146–1Won9432026-11-10
Jev 1.134–2Won9442026-11-10
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.6–0Won9452026-11-10
Cygnet5–2Won9472026-11-11
Winnow-12B6–1Won9482026-11-11
Perplexity Decider Flash v1.26–1Won9492026-11-11
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell5–2Won9502026-11-11
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell5–1Won9512026-11-11
Clef6–2Won9522026-11-11
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell7–2Won9532026-11-11
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–4Draw9542026-11-11
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

KnowLine 4B Gen5.5: results per opponent

Swipe for more columns →

KnowLine 4B Gen5.5 against each opponent, its score first: 5 won, 1 drawn, 11 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI4–5Lost9572026-11-11
celeris-1-decision2–6Lost9582026-11-11
GNUbg weak0–9Lost9592026-11-11
Random8–2Won9602026-11-11
Safest2–5Lost9612026-11-11
d1-ultra0–5Lost9622026-11-11
Jev 1.140–5Lost9632026-11-11
Jev 1.132–5Lost9642026-11-11
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.6–6Draw9652026-11-11
Cygnet6–2Won13712026-11-14
Winnow-12B3–5Lost13722026-11-14
Perplexity Decider Flash v1.24–2Won13732026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell1–6Lost13742026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell1–6Lost13752026-11-14
Clef0–7Lost13762026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell4–2Won13772026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–2Won13782026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

StartLux-Decision-v1.1-2B-Instant: results per opponent

Swipe for more columns →

StartLux-Decision-v1.1-2B-Instant against each opponent, its score first: 3 won, 0 drawn, 14 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI1–5Lost9702026-11-11
celeris-1-decision0–5Lost9712026-11-11
GNUbg weak0–6Lost9722026-11-11
Random5–1Won9732026-11-11
Safest5–3Won9742026-11-11
d1-ultra1–5Lost9752026-11-11
Jev 1.140–5Lost9762026-11-11
Jev 1.131–7Lost9772026-11-11
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.2–5Lost9782026-11-11
Cygnet2–4Lost13792026-11-14
Winnow-12B0–6Lost13802026-11-14
Perplexity Decider Flash v1.24–5Lost13812026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell3–5Lost13822026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell5–2Won13832026-11-14
Clef3–5Lost13842026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell3–5Lost13852026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1–6Lost13862026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

KnowLine-0.8B-Gen1: results per opponent

Swipe for more columns →

KnowLine-0.8B-Gen1 against each opponent, its score first: 1 won, 1 drawn, 15 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI0–7Lost9832026-11-11
celeris-1-decision3–6Lost9842026-11-11
GNUbg weak0–5Lost9852026-11-11
Random6–3Won9862026-11-11
Safest3–4Lost9872026-11-11
d1-ultra2–5Lost9882026-11-11
Jev 1.140–5Lost9892026-11-11
Jev 1.130–5Lost9902026-11-11
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.3–4Lost9912026-11-11
Cygnet1–6Lost13312026-11-14
Winnow-12B2–4Lost13322026-11-14
Perplexity Decider Flash v1.22–6Lost13332026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell0–6Lost13342026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell2–7Lost13352026-11-14
Clef0–5Lost13362026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–5Draw13372026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark2–8Lost13382026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

Gemini Decision 1: results per opponent

Swipe for more columns →

Gemini Decision 1 against each opponent, its score first: 13 won, 0 drawn, 4 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI5–1Won9932026-11-11
celeris-1-decision2–6Lost9942026-11-11
GNUbg weak1–5Lost9952026-11-11
Random5–0Won9962026-11-11
Safest7–1Won9972026-11-11
d1-ultra4–3Won9982026-11-11
Jev 1.146–2Won9992026-11-11
Jev 1.133–5Lost10052026-11-11
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.1–5Lost10132026-11-11
Cygnet6–0Won10532026-11-11
Winnow-12B6–1Won10552026-11-11
Perplexity Decider Flash v1.25–0Won10562026-11-11
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell4–2Won10572026-11-11
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell5–2Won10602026-11-11
Clef6–1Won10612026-11-11
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–3Won10622026-11-11
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–3Won10632026-11-11
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

StartLux-Decision-v1.1-0.8B-Instant: results per opponent

Swipe for more columns →

StartLux-Decision-v1.1-0.8B-Instant against each opponent, its score first: 3 won, 0 drawn, 14 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI2–6Lost10002026-11-11
celeris-1-decision1–5Lost10012026-11-11
GNUbg weak1–6Lost10032026-11-11
Random8–0Won10062026-11-11
Safest1–5Lost10072026-11-11
d1-ultra2–5Lost10092026-11-11
Jev 1.141–5Lost10102026-11-11
Jev 1.131–5Lost10122026-11-11
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.0–5Lost10142026-11-11
Cygnet2–6Lost13392026-11-14
Winnow-12B2–7Lost13402026-11-14
Perplexity Decider Flash v1.23–7Lost13412026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell6–3Won13422026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell2–5Lost13432026-11-14
Clef0–5Lost13442026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell4–2Won13452026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0–5Lost13462026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

OneJev1.6-0.8B: results per opponent

Swipe for more columns →

OneJev1.6-0.8B against each opponent, its score first: 0 won, 0 drawn, 17 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI0–5Lost10412026-11-11
celeris-1-decision0–6Lost10422026-11-11
GNUbg weak0–5Lost10432026-11-11
Random4–5Lost10442026-11-11
Safest1–7Lost10452026-11-11
d1-ultra1–6Lost10462026-11-11
Jev 1.141–8Lost10472026-11-11
Jev 1.131–6Lost10482026-11-11
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.2–5Lost10492026-11-11
Cygnet1–5Lost13472026-11-14
Winnow-12B0–7Lost13482026-11-14
Perplexity Decider Flash v1.21–6Lost13492026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell0–8Lost13502026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell0–7Lost13512026-11-14
Clef1–7Lost13522026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell4–5Lost13532026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0–7Lost13542026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

GLiClass v4 Edge: results per opponent

Swipe for more columns →

GLiClass v4 Edge against each opponent, its score first: 0 won, 0 drawn, 17 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI0–8Lost11412026-11-12
celeris-1-decision1–8Lost11422026-11-12
GNUbg weak0–5Lost11432026-11-12
Random1–5Lost11442026-11-12
Safest0–7Lost11652026-11-12
d1-ultra0–7Lost11662026-11-12
Jev 1.140–6Lost11672026-11-12
Jev 1.130–6Lost11682026-11-12
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.0–7Lost11692026-11-12
Cygnet0–5Lost13632026-11-14
Winnow-12B1–6Lost13642026-11-14
Perplexity Decider Flash v1.20–6Lost13652026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell1–5Lost13662026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell0–7Lost13672026-11-14
Clef0–8Lost13682026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell0–7Lost13692026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0–5Lost13702026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

KnowLine-2B-Gen1: results per opponent

Swipe for more columns →

KnowLine-2B-Gen1 against each opponent, its score first: 3 won, 0 drawn, 14 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI3–5Lost11812026-11-12
celeris-1-decision2–5Lost11822026-11-12
GNUbg weak0–6Lost11832026-11-12
Random5–2Won11842026-11-12
Safest5–3Won11852026-11-12
d1-ultra1–5Lost11862026-11-12
Jev 1.143–5Lost11872026-11-12
Jev 1.132–5Lost11882026-11-12
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.2–6Lost11892026-11-12
Cygnet1–5Lost13552026-11-14
Winnow-12B2–5Lost13562026-11-14
Perplexity Decider Flash v1.23–4Lost13572026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell1–6Lost13582026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell2–5Lost13592026-11-14
Clef5–2Won13602026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell3–6Lost13612026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark3–4Lost13622026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

GLiDE 2B: results per opponent

Swipe for more columns →

GLiDE 2B against each opponent, its score first: 1 won, 0 drawn, 16 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI1–7Lost11912026-11-12
celeris-1-decision2–4Lost11922026-11-12
GNUbg weak0–6Lost11932026-11-12
Random5–0Won11942026-11-12
Safest1–6Lost11952026-11-12
d1-ultra0–7Lost11962026-11-12
Jev 1.140–5Lost11972026-11-12
Jev 1.130–5Lost11982026-11-12
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.1–6Lost11992026-11-12
Cygnet2–6Lost12982026-11-13
Winnow-12B4–6Lost12992026-11-14
Perplexity Decider Flash v1.21–5Lost13002026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell1–5Lost13012026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell2–5Lost13022026-11-14
Clef0–6Lost13032026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell3–4Lost13042026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1–6Lost13052026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

GLiDE 0.8B: results per opponent

Swipe for more columns →

GLiDE 0.8B against each opponent, its score first: 2 won, 1 drawn, 14 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI0–6Lost12012026-11-12
celeris-1-decision0–7Lost12022026-11-12
GNUbg weak1–7Lost12032026-11-12
Random6–5Won12042026-11-12
Safest4–4Draw12052026-11-12
d1-ultra2–4Lost12062026-11-12
Jev 1.141–7Lost12072026-11-12
Jev 1.130–5Lost12082026-11-12
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.4–5Lost12092026-11-12
Cygnet4–3Won13152026-11-14
Winnow-12B2–6Lost13162026-11-14
Perplexity Decider Flash v1.23–4Lost13172026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell1–6Lost13182026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell1–6Lost13192026-11-14
Clef1–6Lost13202026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell2–5Lost13212026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1–6Lost13222026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

GLiDE 0.4B: results per opponent

Swipe for more columns →

GLiDE 0.4B against each opponent, its score first: 1 won, 0 drawn, 16 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI0–7Lost12112026-11-12
celeris-1-decision0–8Lost12122026-11-12
GNUbg weak0–6Lost12132026-11-12
Random7–3Won12142026-11-12
Safest0–6Lost12152026-11-12
d1-ultra0–8Lost12162026-11-12
Jev 1.140–7Lost12172026-11-12
Jev 1.130–7Lost12182026-11-12
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.0–6Lost12192026-11-12
Cygnet1–5Lost13232026-11-14
Winnow-12B0–6Lost13242026-11-14
Perplexity Decider Flash v1.21–5Lost13252026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell0–9Lost13262026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell0–8Lost13272026-11-14
Clef0–5Lost13282026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell1–6Lost13292026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0–5Lost13302026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

Sieve 31B v3: results per opponent

Swipe for more columns →

Sieve 31B v3 against each opponent, its score first: 15 won, 0 drawn, 2 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI4–3Won12212026-11-12
celeris-1-decision6–2Won12222026-11-12
GNUbg weak4–3Won12232026-11-12
Random9–0Won12242026-11-12
Safest5–1Won12252026-11-12
d1-ultra3–4Lost12262026-11-12
Jev 1.144–3Won12272026-11-12
Jev 1.135–1Won12282026-11-12
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.3–5Lost12292026-11-12
Cygnet5–2Won14112026-11-14
Winnow-12B5–0Won14122026-11-14
Perplexity Decider Flash v1.25–3Won14132026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell5–2Won14142026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell6–0Won14152026-11-14
Clef5–1Won14162026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–1Won14172026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark5–1Won14182026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

Turion Jev 31B: results per opponent

Swipe for more columns →

Turion Jev 31B against each opponent, its score first: 13 won, 1 drawn, 3 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI6–0Won12412026-11-13
celeris-1-decision5–3Won12422026-11-13
GNUbg weak6–0Won12432026-11-13
Random7–0Won12442026-11-13
Safest5–2Won12452026-11-13
d1-ultra2–5Lost12462026-11-13
Jev 1.145–2Won12472026-11-13
Jev 1.133–4Lost12482026-11-13
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.5–5Draw12492026-11-13
Cygnet4–3Won14032026-11-14
Winnow-12B7–0Won14042026-11-14
Perplexity Decider Flash v1.25–0Won14052026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell5–4Won14062026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell6–1Won14072026-11-14
Clef5–6Lost14082026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–2Won14092026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark6–3Won14102026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

StartLux-Decision-v1.1-4B-Instant: results per opponent

Swipe for more columns →

StartLux-Decision-v1.1-4B-Instant against each opponent, its score first: 10 won, 0 drawn, 7 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI4–2Won12512026-11-13
celeris-1-decision4–2Won12522026-11-13
GNUbg weak1–6Lost12532026-11-13
Random6–0Won12542026-11-13
Safest5–0Won12552026-11-13
d1-ultra1–5Lost12562026-11-13
Jev 1.146–3Won12572026-11-13
Jev 1.137–5Won12582026-11-13
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.5–3Won12592026-11-13
Cygnet5–2Won13872026-11-14
Winnow-12B1–6Lost13882026-11-14
Perplexity Decider Flash v1.26–1Won13892026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell1–6Lost13902026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell2–6Lost13912026-11-14
Clef3–4Lost13922026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–3Won13932026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–6Lost13942026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

Cerno 9B v1.2: results per opponent

Swipe for more columns →

Cerno 9B v1.2 against each opponent, its score first: 6 won, 1 drawn, 10 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI4–6Lost12612026-11-13
celeris-1-decision2–6Lost12622026-11-13
GNUbg weak1–7Lost12632026-11-13
Random8–0Won12642026-11-13
Safest6–1Won12652026-11-13
d1-ultra2–6Lost12662026-11-13
Jev 1.143–6Lost12672026-11-13
Jev 1.131–6Lost12682026-11-13
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.3–5Lost12692026-11-13
Cygnet5–5Draw13952026-11-14
Winnow-12B2–4Lost13962026-11-14
Perplexity Decider Flash v1.26–1Won13972026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell5–4Won13982026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell4–5Lost13992026-11-14
Clef5–4Won14002026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell6–1Won14012026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark4–6Lost14022026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

d1-2BNVIDIA RTX PRO 6000 Blackwell: results per opponent

Swipe for more columns →

d1-2BNVIDIA RTX PRO 6000 Blackwell against each opponent, its score first: 5 won, 0 drawn, 12 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI1–7Lost14192026-11-14
celeris-1-decision2–5Lost14202026-11-14
GNUbg weak3–4Lost14212026-11-14
Random6–1Won14222026-11-14
Safest5–2Won14232026-11-14
d1-ultra1–6Lost14242026-11-14
Jev 1.142–5Lost14252026-11-14
Jev 1.132–5Lost14262026-11-14
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.1–5Lost14272026-11-14
Cygnet5–2Won14292026-11-14
Winnow-12B2–6Lost14302026-11-14
Perplexity Decider Flash v1.24–2Won14312026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell4–5Lost14322026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell2–6Lost14332026-11-14
Clef2–5Lost14342026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–2Won14352026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1–6Lost14362026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

d1-Omni-2BNVIDIA RTX PRO 6000 Blackwell: results per opponent

Swipe for more columns →

d1-Omni-2BNVIDIA RTX PRO 6000 Blackwell against each opponent, its score first: 9 won, 0 drawn, 8 lost; 2 outside the requested scope (Ümit, 12 Nov), not played, not a loss; none counts in a standings table
OpponentScoreResult#Finished
GLiDEAPI2–4Lost14372026-11-14
celeris-1-decision6–3Won14382026-11-14
GNUbg weak1–6Lost14392026-11-14
Random5–1Won14402026-11-14
Safest1–5Lost14412026-11-14
d1-ultra6–1Won14422026-11-14
Jev 1.145–2Won14432026-11-14
Jev 1.136–1Won14442026-11-14
SubQ Snap PreviewThe opponent was requested as SubQ Snap Preview; under that name the API returned answers identical to SubQ Snap 1.1 on 1,000 paired positions, so the opponent's model identity is not verified.3–6Lost14452026-11-14
Cygnet0–5Lost14472026-11-14
Winnow-12B0–6Lost14482026-11-14
Perplexity Decider Flash v1.25–2Won14492026-11-14
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell7–2Won14502026-11-14
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell1–6Lost14512026-11-14
Clef7–3Won14522026-11-14
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell5–2Won14532026-11-14
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0–5Lost14542026-11-14
GPT-6 LunaDecisions APINo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)N/A
Mercury DecideNo scoreOutside the requested scope (Ümit, 12 Nov); not played (opponent no longer served on OpenRouter)N/A

Error rate and risk judgement: 95% intervals, whole games resampled. Ability d′ and caution c at the 0.001 margin: rates corrected by adding 0.5 to the counts and 1 to the totals; 95% intervals, blocks of two seat-swapped games resampled. Cut off: answers that stopped at an output limit the request sent (finish reason length, stop reason max_tokens, finishReason MAX_TOKENS, incomplete max_output_tokens), in the stored response or error. Effort: the reasoning effort the requests sent. Engine's best play: the share of its scored decisions that chose the engine's best play (equity loss under the best-play threshold). Time: median request seconds per decision (waits between retries not included). Cost per 1,000 decisions: the recorded tokens at the price shown, as labelled there (list or billed price; for a setup on our GPUs the hosted model's list price, not billed). Every match of every stage: Matches.

Risk judgement

Risk calls per setup: was each risk the right one, and was safe right when it played safe? 100 = every call right, 50 = always safe; the line on each bar is its 95% interval (games resampled) · higher is better

Leagues played different positions; close scores across leagues need not mean a real difference.

  1. Decision models
  2. d1-ultra · League 264.0 | 61.7–66.3
  3. GLiDEAPI · League 260.4 | 58.3–62.6
  4. Jev 1.13 · League 159.5 | 57.4–61.6
  5. Jev 1.14 · League 259.3 | 57.1–61.4
  6. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell · League 158.8 | 56.7–60.8
  7. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell · League 157.9 | 56.0–59.9
  8. celeris-1-decision · League 157.8 | 55.9–59.7
  9. Clef · League 257.5 | 55.4–59.8
  10. Cygnet · League 157.1 | 55.3–58.8
  11. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 156.9 | 54.9–58.8
  12. SubQ Snap Preview · League 156.5 | 53.8–59.1
  13. GPT-6 LunaDecisions API · League 155.7 | 53.8–57.6
  14. Winnow-12B · League 155.2 | 52.9–57.2
  15. Mercury Decide · League 153.4 | 51.5–55.2
  16. Perplexity Decider Flash v1.2 · League 149.1 | 46.5–51.6
  17. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell · League 148.5 | 46.1–50.9
  18. Chat models
  19. GPT-6 Astralow · League 375.8 | 73.1–78.5
  20. GPT-6 Astramedium · League 374.7 | 71.9–77.3
  21. Gemini 4 Argonhigh · League 374.1 | 72.0–76.1
  22. GPT-6.1 Sollow · League 372.7 | 69.9–75.5
  23. Claude Fable 5.1high · League 372.3 | 69.8–74.7
  24. GPT-6.1 Solhigh · League 372.2 | 69.5–74.9
  25. Claude Opus 5.5high · League 370.0 | 67.3–72.6
  26. Gemini 4 Argonlow · League 369.2 | 66.2–72.3
  27. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 269.0 | 67.3–70.9
  28. Grok 4.7high · League 368.5 | 66.1–71.1
  29. DeepSeek-V4.1-Flashlow · API · 1017 · League 268.5 | 66.4–70.6
  30. Claude Sonnet 5.5high · League 268.2 | 66.3–70.0
  31. Claude Fable 5.1low · League 267.8 | 65.6–69.9
  32. Grok 4.7low · League 267.7 | 66.0–69.5
  33. Claude Opus 5.5low · League 366.9 | 64.7–69.2
  34. Gemini 3.8 Flashhigh · League 266.9 | 65.2–68.5
  35. Kimi K3.1 Prooff · League 365.9 | 63.2–68.9
  36. Gemini 3.8 Flashlow · League 265.1 | 63.5–66.8
  37. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 264.5 | 62.9–66.2
  38. Claude Sonnet 5.5low · League 263.5 | 61.6–65.4
  39. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 262.6 | 60.4–65.0
  40. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 262.0 | 60.1–63.9
  41. GLM-5.3low · League 261.7 | 59.7–63.8
  42. Qwen3.8 Flashoff · League 259.5 | 57.2–61.7
  43. DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 159.3 | 57.1–61.5
  44. DeepSeek-V4.1-Flashoff · API · 1017 · League 258.5 | 56.6–60.5
  45. Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 157.6 | 56.1–59.0
  46. LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 152.4 | 50.3–54.5
  47. LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell · League 145.7 | 43.2–48.1
  48. Programs: reference rows, never ranked
  49. GNUbg weak · League 167.1 | 63.8–70.3
  50. GNUbg weak · League 262.3 | 59.5–64.8
  51. GNUbg weak · League 362.1 | 58.2–65.9
  52. Safest · League 150.0 | 50.0–50.0
  53. Safest (computed) · League 250.0 | 50.0–50.0
  54. Safest (computed) · League 350.0 | 50.0–50.0
  55. Random (computed) · League 327.6 | 27.1–28.0
  56. Random (computed) · League 227.0 | 26.7–27.2
  57. Random · League 124.9 | 22.3–27.4
  1. Decision models
  2. Jev 1.1359.5 | 57.4–61.6
  3. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell58.8 | 56.7–60.8
  4. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell57.9 | 56.0–59.9
  5. celeris-1-decision57.8 | 55.9–59.7
  6. Cygnet57.1 | 55.3–58.8
  7. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark56.9 | 54.9–58.8
  8. SubQ Snap Preview56.5 | 53.8–59.1
  9. GPT-6 LunaDecisions API55.7 | 53.8–57.6
  10. Winnow-12B55.2 | 52.9–57.2
  11. Mercury Decide53.4 | 51.5–55.2
  12. Perplexity Decider Flash v1.249.1 | 46.5–51.6
  13. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell48.5 | 46.1–50.9
  14. Chat models
  15. DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release59.3 | 57.1–61.5
  16. Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark57.6 | 56.1–59.0
  17. LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark52.4 | 50.3–54.5
  18. LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell45.7 | 43.2–48.1
  19. Programs: reference rows, never ranked
  20. GNUbg weak67.1 | 63.8–70.3
  21. Safest50.0 | 50.0–50.0
  22. Random24.9 | 22.3–27.4
  1. Decision models
  2. d1-ultra64.0 | 61.7–66.3
  3. GLiDEAPI60.4 | 58.3–62.6
  4. Jev 1.1459.3 | 57.1–61.4
  5. Clef57.5 | 55.4–59.8
  6. Chat models
  7. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release69.0 | 67.3–70.9
  8. DeepSeek-V4.1-Flashlow · API · 101768.5 | 66.4–70.6
  9. Claude Sonnet 5.5high68.2 | 66.3–70.0
  10. Claude Fable 5.1low67.8 | 65.6–69.9
  11. Grok 4.7low67.7 | 66.0–69.5
  12. Gemini 3.8 Flashhigh66.9 | 65.2–68.5
  13. Gemini 3.8 Flashlow65.1 | 63.5–66.8
  14. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark64.5 | 62.9–66.2
  15. Claude Sonnet 5.5low63.5 | 61.6–65.4
  16. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark62.6 | 60.4–65.0
  17. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark62.0 | 60.1–63.9
  18. GLM-5.3low61.7 | 59.7–63.8
  19. Qwen3.8 Flashoff59.5 | 57.2–61.7
  20. DeepSeek-V4.1-Flashoff · API · 101758.5 | 56.6–60.5
  21. Programs: reference rows, never ranked
  22. GNUbg weak62.3 | 59.5–64.8
  23. Safest (computed)50.0 | 50.0–50.0
  24. Random (computed)27.0 | 26.7–27.2
  1. Chat models
  2. GPT-6 Astralow75.8 | 73.1–78.5
  3. GPT-6 Astramedium74.7 | 71.9–77.3
  4. Gemini 4 Argonhigh74.1 | 72.0–76.1
  5. GPT-6.1 Sollow72.7 | 69.9–75.5
  6. Claude Fable 5.1high72.3 | 69.8–74.7
  7. GPT-6.1 Solhigh72.2 | 69.5–74.9
  8. Claude Opus 5.5high70.0 | 67.3–72.6
  9. Gemini 4 Argonlow69.2 | 66.2–72.3
  10. Grok 4.7high68.5 | 66.1–71.1
  11. Claude Opus 5.5low66.9 | 64.7–69.2
  12. Kimi K3.1 Prooff65.9 | 63.2–68.9
  13. Programs: reference rows, never ranked
  14. GNUbg weak62.1 | 58.2–65.9
  15. Safest (computed)50.0 | 50.0–50.0
  16. Random (computed)27.6 | 27.1–28.0

Analysis pending Shared positions: Analysis pending.

OpenAIGoogleAnthropicxAIDeepSeekMoonshotLiquid AIAlibabaInceptionCloudflareZ.aiFastinoPerplexitySubquadraticCelerisEldanRingTypeSafeZefan CaiBlockbrain Labsprogram (reference row)95% interval 50 = always safe

Ranks show observed scores; small gaps need not be meaningful. More caveats: Method

Error rate

How much equity each move gives away, judged by the engine, not by wins: error rate (PR) with its 95% interval as the line on each bar, log scale · lower is better

Leagues played different positions; close scores across leagues need not mean a real difference.

  1. Decision models
  2. d1-ultra · League 230.3 | 27.3–33.4
  3. Jev 1.14 · League 231.7 | 28.3–35.7
  4. GLiDEAPI · League 233.2 | 30.2–36.4
  5. Jev 1.13 · League 134.4 | 31.0–38.0
  6. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell · League 135.0 | 31.6–39.0
  7. celeris-1-decision · League 135.6 | 32.0–39.4
  8. Clef · League 236.5 | 33.5–39.6
  9. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell · League 136.9 | 33.9–40.1
  10. Cygnet · League 138.6 | 35.4–41.9
  11. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 138.8 | 35.3–42.5
  12. GPT-6 LunaDecisions API · League 143.3 | 39.7–46.9
  13. SubQ Snap Preview · League 143.9 | 39.3–48.8
  14. Mercury Decide · League 147.9 | 43.3–52.9
  15. Winnow-12B · League 149.7 | 45.9–53.6
  16. Perplexity Decider Flash v1.2 · League 159.6 | 54.9–64.5
  17. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell · League 165.6 | 59.7–71.9
  18. Chat models
  19. GPT-6.1 Solhigh · League 39.9 | 8.6–11.3
  20. GPT-6 Astramedium · League 311.1 | 9.5–12.9
  21. GPT-6 Astralow · League 311.8 | 9.5–14.2
  22. GPT-6.1 Sollow · League 312.1 | 10.5–13.7
  23. Gemini 4 Argonhigh · League 312.6 | 10.5–14.8
  24. Claude Fable 5.1high · League 314.0 | 11.5–16.3
  25. Claude Opus 5.5high · League 314.4 | 12.1–17.0
  26. Gemini 3.8 Flashhigh · League 214.9 | 12.9–17.0
  27. Claude Opus 5.5low · League 315.4 | 13.1–18.0
  28. Gemini 3.8 Flashlow · League 216.3 | 14.5–18.2
  29. Gemini 4 Argonlow · League 316.5 | 13.7–19.5
  30. Claude Sonnet 5.5high · League 216.7 | 14.9–18.7
  31. DeepSeek-V4.1-Flashlow · API · 1017 · League 217.1 | 15.1–19.2
  32. Claude Fable 5.1low · League 217.9 | 15.9–20.1
  33. Grok 4.7high · League 318.8 | 15.9–21.8
  34. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 219.0 | 16.6–21.3
  35. Grok 4.7low · League 219.1 | 16.9–21.4
  36. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 219.6 | 17.5–21.7
  37. Claude Sonnet 5.5low · League 220.1 | 17.8–22.4
  38. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 222.5 | 19.9–25.3
  39. Kimi K3.1 Prooff · League 328.0 | 24.5–31.8
  40. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 229.0 | 26.5–31.8
  41. GLM-5.3low · League 230.9 | 27.8–34.0
  42. Qwen3.8 Flashoff · League 233.6 | 30.8–36.7
  43. DeepSeek-V4.1-Flashoff · API · 1017 · League 235.2 | 31.8–38.8
  44. DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 135.6 | 31.7–39.7
  45. Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 145.0 | 40.7–49.9
  46. LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 147.7 | 44.1–51.2
  47. LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell · League 161.5 | 56.3–66.8
  48. Programs: reference rows, never ranked
  49. GNUbg weak · League 121.4 | 18.8–24.3
  50. GNUbg weak · League 227.0 | 24.1–29.8
  51. GNUbg weak · League 328.1 | 25.0–31.4
  52. Safest · League 151.6 | 47.4–56.2
  53. Random · League 1121.1 | 113.3–128.5
  1. Decision models
  2. Jev 1.1334.4 | 31.0–38.0
  3. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell35.0 | 31.6–39.0
  4. celeris-1-decision35.6 | 32.0–39.4
  5. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell36.9 | 33.9–40.1
  6. Cygnet38.6 | 35.4–41.9
  7. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark38.8 | 35.3–42.5
  8. GPT-6 LunaDecisions API43.3 | 39.7–46.9
  9. SubQ Snap Preview43.9 | 39.3–48.8
  10. Mercury Decide47.9 | 43.3–52.9
  11. Winnow-12B49.7 | 45.9–53.6
  12. Perplexity Decider Flash v1.259.6 | 54.9–64.5
  13. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell65.6 | 59.7–71.9
  14. Chat models
  15. DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release35.6 | 31.7–39.7
  16. Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark45.0 | 40.7–49.9
  17. LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark47.7 | 44.1–51.2
  18. LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell61.5 | 56.3–66.8
  19. Programs: reference rows, never ranked
  20. GNUbg weak21.4 | 18.8–24.3
  21. Safest51.6 | 47.4–56.2
  22. Random121.1 | 113.3–128.5
  1. Decision models
  2. d1-ultra30.3 | 27.3–33.4
  3. Jev 1.1431.7 | 28.3–35.7
  4. GLiDEAPI33.2 | 30.2–36.4
  5. Clef36.5 | 33.5–39.6
  6. Chat models
  7. Gemini 3.8 Flashhigh14.9 | 12.9–17.0
  8. Gemini 3.8 Flashlow16.3 | 14.5–18.2
  9. Claude Sonnet 5.5high16.7 | 14.9–18.7
  10. DeepSeek-V4.1-Flashlow · API · 101717.1 | 15.1–19.2
  11. Claude Fable 5.1low17.9 | 15.9–20.1
  12. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release19.0 | 16.6–21.3
  13. Grok 4.7low19.1 | 16.9–21.4
  14. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark19.6 | 17.5–21.7
  15. Claude Sonnet 5.5low20.1 | 17.8–22.4
  16. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark22.5 | 19.9–25.3
  17. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark29.0 | 26.5–31.8
  18. GLM-5.3low30.9 | 27.8–34.0
  19. Qwen3.8 Flashoff33.6 | 30.8–36.7
  20. DeepSeek-V4.1-Flashoff · API · 101735.2 | 31.8–38.8
  21. Programs: reference rows, never ranked
  22. GNUbg weak27.0 | 24.1–29.8
  1. Chat models
  2. GPT-6.1 Solhigh9.9 | 8.6–11.3
  3. GPT-6 Astramedium11.1 | 9.5–12.9
  4. GPT-6 Astralow11.8 | 9.5–14.2
  5. GPT-6.1 Sollow12.1 | 10.5–13.7
  6. Gemini 4 Argonhigh12.6 | 10.5–14.8
  7. Claude Fable 5.1high14.0 | 11.5–16.3
  8. Claude Opus 5.5high14.4 | 12.1–17.0
  9. Claude Opus 5.5low15.4 | 13.1–18.0
  10. Gemini 4 Argonlow16.5 | 13.7–19.5
  11. Grok 4.7high18.8 | 15.9–21.8
  12. Kimi K3.1 Prooff28.0 | 24.5–31.8
  13. Programs: reference rows, never ranked
  14. GNUbg weak28.1 | 25.0–31.4

Analysis pending Shared positions: Analysis pending.

OpenAIGoogleAnthropicxAIDeepSeekMoonshotLiquid AIAlibabaInceptionCloudflareZ.aiFastinoPerplexitySubquadraticCelerisEldanRingTypeSafeZefan CaiBlockbrain Labsprogram (reference row)95% interval a program's error rate (reference row, named above the axis)

Setups played different positions; league comparisons are not paired. More caveats: Method

Cost

Dollars per 1,000 decisionsper 1M decisions, log scale · lower is better

Leagues played different positions; close scores across leagues need not mean a real difference.

  1. Decision models
  2. Mercury Decide · League 1$0.071$0.071
  3. d1-ultra · League 2$0.145$0.145
  4. celeris-1-decision · League 1$0.150$0.150
  5. GLiDEAPI · League 2$0.153$0.153
  6. Jev 1.13 · League 1$0.166$0.166
  7. Jev 1.14 · League 2$0.173$0.173
  8. SubQ Snap Preview · League 1$0.193$0.193
  9. GPT-6 LunaDecisions API · League 1$0.335$0.335
  10. Clef · League 2$5.736$5.736 estimate
  11. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 1$6.593$6.593 estimate
  12. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell · League 1$7.185$7.185 estimate
  13. Winnow-12B · League 1$9.615$9.615 estimate
  14. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell · League 1$10.301$10.301 estimate
  15. Cygnet · League 1$11.168$11.168 estimate
  16. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell · League 1$12.227$12.227 estimate
  17. Perplexity Decider Flash v1.2 · League 1$12.267$12.267 estimate
  18. Chat models
  19. DeepSeek-V4.1-Flashoff · API · 1017 · League 2$0.244$0.244
  20. Qwen3.8 Flashoff · League 2$0.284$0.284
  21. DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 1$0.728$0.728 estimate
  22. GLM-5.3low · League 2$2.917$2.917
  23. Kimi K3.1 Prooff · League 3$3.982$3.982
  24. Gemini 3.8 Flashlow · League 2$4.123$4.123
  25. Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 1$4.551$4.551 estimate
  26. DeepSeek-V4.1-Flashlow · API · 1017 · League 2$4.869$4.869
  27. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 2$8.215$8.215 estimate
  28. LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell · League 1$9.132$9.132 estimate
  29. Gemini 4 Argonlow · League 3$10.204$10.204
  30. GPT-6.1 Sollow · League 3$10.228$10.228
  31. Claude Sonnet 5.5low · League 2$13.541$13.541
  32. GPT-6.1 Solhigh · League 3$15.573$15.573
  33. Claude Sonnet 5.5high · League 2$19.847$19.847
  34. Gemini 3.8 Flashhigh · League 2$22.352$22.352
  35. Claude Opus 5.5low · League 3$29.062$29.062
  36. Grok 4.7low · League 2$32.813$32.813
  37. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 2$34.943$34.943 estimate
  38. Claude Opus 5.5high · League 3$36.400$36.400
  39. LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 1$38.716$38.716 estimate
  40. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 2$39.023$39.023 estimate
  41. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 2$47.092$47.092 estimate
  42. GPT-6 Astralow · League 3$49.737$49.737
  43. GPT-6 Astramedium · League 3$50.268$50.268
  44. Claude Fable 5.1low · League 2$62.613$62.613
  45. Gemini 4 Argonhigh · League 3$64.401$64.401
  46. Grok 4.7high · League 3$117.379$117.379
  47. Claude Fable 5.1high · League 3$136.268$136.268
  1. Decision models
  2. Mercury Decide$0.071$0.071
  3. celeris-1-decision$0.150$0.150
  4. Jev 1.13$0.166$0.166
  5. SubQ Snap Preview$0.193$0.193
  6. GPT-6 LunaDecisions API$0.335$0.335
  7. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark$6.593$6.593 estimate
  8. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell$7.185$7.185 estimate
  9. Winnow-12B$9.615$9.615 estimate
  10. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell$10.301$10.301 estimate
  11. Cygnet$11.168$11.168 estimate
  12. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell$12.227$12.227 estimate
  13. Perplexity Decider Flash v1.2$12.267$12.267 estimate
  14. Chat models
  15. DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release$0.728$0.728 estimate
  16. Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark$4.551$4.551 estimate
  17. LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell$9.132$9.132 estimate
  18. LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark$38.716$38.716 estimate
  1. Decision models
  2. d1-ultra$0.145$0.145
  3. GLiDEAPI$0.153$0.153
  4. Jev 1.14$0.173$0.173
  5. Clef$5.736$5.736 estimate
  6. Chat models
  7. DeepSeek-V4.1-Flashoff · API · 1017$0.244$0.244
  8. Qwen3.8 Flashoff$0.284$0.284
  9. GLM-5.3low$2.917$2.917
  10. Gemini 3.8 Flashlow$4.123$4.123
  11. DeepSeek-V4.1-Flashlow · API · 1017$4.869$4.869
  12. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark$8.215$8.215 estimate
  13. Claude Sonnet 5.5low$13.541$13.541
  14. Claude Sonnet 5.5high$19.847$19.847
  15. Gemini 3.8 Flashhigh$22.352$22.352
  16. Grok 4.7low$32.813$32.813
  17. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark$34.943$34.943 estimate
  18. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark$39.023$39.023 estimate
  19. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release$47.092$47.092 estimate
  20. Claude Fable 5.1low$62.613$62.613
  1. Chat models
  2. Kimi K3.1 Prooff$3.982$3.982
  3. Gemini 4 Argonlow$10.204$10.204
  4. GPT-6.1 Sollow$10.228$10.228
  5. GPT-6.1 Solhigh$15.573$15.573
  6. Claude Opus 5.5low$29.062$29.062
  7. Claude Opus 5.5high$36.400$36.400
  8. GPT-6 Astralow$49.737$49.737
  9. GPT-6 Astramedium$50.268$50.268
  10. Gemini 4 Argonhigh$64.401$64.401
  11. Grok 4.7high$117.379$117.379
  12. Claude Fable 5.1high$136.268$136.268

Analysis pending Shared positions: Analysis pending.

Programs, Not measured: GNUbg weak (League 1, 2, 3); Random (League 1); Safest (League 1).

OpenAIGoogleAnthropicxAIDeepSeekMoonshotLiquid AIAlibabaInceptionCloudflareZ.aiFastinoPerplexitySubquadraticCelerisEldanRingTypeSafeZefan CaiBlockbrain Labsprogram (reference row)

Setups on our own machines: an estimate, GPU time, at most (wall-clock time × hourly rate, an upper bound); each setup's price label and how its cost was counted are in its row's tooltip and in the Cost tab. More caveats: Method

Time

Median seconds per decision (P90 beside it), log scale · lower is better

Leagues played different positions; close scores across leagues need not mean a real difference.

  1. Decision models
  2. celeris-1-decision · League 10.17 · 0.33
  3. Jev 1.14 · League 20.19 · 0.24
  4. Jev 1.13 · League 10.22 · 0.26
  5. d1-ultra · League 20.23 · 0.41
  6. GPT-6 LunaDecisions API · League 10.23 · 0.35
  7. SubQ Snap Preview · League 10.24 · 0.32
  8. Cygnet · League 10.37 · 2.96
  9. GLiDEAPI · League 20.48 · 0.79
  10. Mercury Decide · League 10.48 · 0.77
  11. Perplexity Decider Flash v1.2 · League 10.56 · 0.96
  12. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell · League 10.94 · 2.73
  13. Clef · League 21.13 · 5.41
  14. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell · League 11.41 · 2.75
  15. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell · League 11.57 · 3.38
  16. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 11.97 · 7.48
  17. Winnow-12B · League 12.46 · 5.45
  18. Chat models
  19. LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell · League 10.14 · 0.36
  20. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 20.31 · 4.03
  21. DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 10.48 · 1.63
  22. DeepSeek-V4.1-Flashoff · API · 1017 · League 20.84 · 1.39
  23. Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 11.48 · 2.71
  24. Claude Sonnet 5.5low · League 21.51 · 7.37
  25. Claude Fable 5.1low · League 21.54 · 6.34
  26. Qwen3.8 Flashoff · League 21.70 · 7.47
  27. GLM-5.3low · League 21.75 · 7.48
  28. Kimi K3.1 Prooff · League 32.11 · 7.23
  29. Gemini 3.8 Flashlow · League 22.16 · 6.87
  30. Gemini 4 Argonlow · League 33.82 · 12.2
  31. Claude Opus 5.5low · League 33.92 · 10.6
  32. GPT-6 Astralow · League 33.94 · 14.0
  33. GPT-6.1 Sollow · League 34.33 · 14.3
  34. GPT-6 Astramedium · League 34.40 · 15.1
  35. Claude Opus 5.5high · League 36.35 · 18.5
  36. Claude Sonnet 5.5high · League 26.63 · 17.6
  37. Gemini 3.8 Flashhigh · League 29.59 · 44.4
  38. Claude Fable 5.1high · League 312.7 · 47.1
  39. GPT-6.1 Solhigh · League 313.1 · 56.5
  40. DeepSeek-V4.1-Flashlow · API · 1017 · League 222.3 · 107.6
  41. Grok 4.7low · League 224.3 · 110.2
  42. Gemini 4 Argonhigh · League 327.5 · 148.6
  43. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release · League 243.8 · 173.1
  44. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 2120.6 · 394.3
  45. Grok 4.7high · League 3122.3 · 588.6
  46. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 2171.1 · 566.7
  47. LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 1324.8 · 668.6
  48. Programs: reference rows, never ranked
  49. Random · League 10.00 · 0.00
  50. Safest · League 10.00 · 0.00
  51. GNUbg weak · League 10.01 · 0.01
  52. GNUbg weak · League 20.01 · 0.01
  53. GNUbg weak · League 30.01 · 0.01
  1. Decision models
  2. Jev 1.140.19 · 0.24
  3. d1-ultra0.23 · 0.41
  4. GLiDEAPI0.48 · 0.79
  5. Clef1.13 · 5.41
  6. Chat models
  7. Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0.31 · 4.03
  8. DeepSeek-V4.1-Flashoff · API · 10170.84 · 1.39
  9. Claude Sonnet 5.5low1.51 · 7.37
  10. Claude Fable 5.1low1.54 · 6.34
  11. Qwen3.8 Flashoff1.70 · 7.47
  12. GLM-5.3low1.75 · 7.48
  13. Gemini 3.8 Flashlow2.16 · 6.87
  14. Claude Sonnet 5.5high6.63 · 17.6
  15. Gemini 3.8 Flashhigh9.59 · 44.4
  16. DeepSeek-V4.1-Flashlow · API · 101722.3 · 107.6
  17. Grok 4.7low24.3 · 110.2
  18. DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release43.8 · 173.1
  19. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark120.6 · 394.3
  20. Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark171.1 · 566.7
  21. Programs: reference rows, never ranked
  22. GNUbg weak0.01 · 0.01
  1. Chat models
  2. Kimi K3.1 Prooff2.11 · 7.23
  3. Gemini 4 Argonlow3.82 · 12.2
  4. Claude Opus 5.5low3.92 · 10.6
  5. GPT-6 Astralow3.94 · 14.0
  6. GPT-6.1 Sollow4.33 · 14.3
  7. GPT-6 Astramedium4.40 · 15.1
  8. Claude Opus 5.5high6.35 · 18.5
  9. Claude Fable 5.1high12.7 · 47.1
  10. GPT-6.1 Solhigh13.1 · 56.5
  11. Gemini 4 Argonhigh27.5 · 148.6
  12. Grok 4.7high122.3 · 588.6
  13. Programs: reference rows, never ranked
  14. GNUbg weak0.01 · 0.01

Analysis pending Shared positions: Analysis pending.

OpenAIGoogleAnthropicxAIDeepSeekMoonshotLiquid AIAlibabaInceptionCloudflareZ.aiFastinoPerplexitySubquadraticCelerisEldanRingTypeSafeZefan CaiBlockbrain Labsprogram (reference row)

Setups played different positions; league comparisons are not paired. More caveats: Method

Every number, setup by setup

Leagues played different positions; close scores across leagues need not mean a real difference.

Swipe for more columns →

All leagues: risk calls at the 0.001 margin, all phases; shares with count / denominator, risk judgement with its 95% interval (games resampled)
Model and settingLeagueRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksError rate by risk-call category (PR)Decisions with a trade-offBig risks foundBig needless risksExtra hitting rolls: right / wrong and needlessConfidence on the call: ranking score for right riskAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
GPT-6 AstralowchatLeague 375.873.1 to 78.557.2%226 / 39594.4%486 / 51584.0%226 / 26939.2%155 / 3955.6%29 / 5153.5%14 / 39511.8needless risks 0.64 · missed risks 6.63 · wrong risks 0.69 · right risks 1.81 · right safety 1.03 · No trade-off 1.00910Risk pays 395 · Safety pays 515 · No trade-off 34168.9%153 / 2221.6%8 / 5159.7 / Too few decisions226 / 43 decisionsN/A1.851.63 to 2.14 · risks taken 240 / 395 Risk pays · needless 29 / 515 Safety pays+0.65+0.57 to +0.76 · cautious94.2%226 / 240unstable: fewer than 200 needless and missed risks behind a mean cost −0.21β 0.68 · needless risks lost 0.056 on average (29) · missed risks 0.107 (155)
GPT-6 AstramediumchatLeague 374.771.9 to 77.354.9%245 / 44694.4%542 / 57484.8%245 / 28942.4%189 / 4465.6%32 / 5742.7%12 / 44611.1needless risks 0.92 · missed risks 6.28 · wrong risks 0.34 · right risks 1.58 · right safety 1.31 · No trade-off 0.671,020Risk pays 446 · Safety pays 574 · No trade-off 34270.5%170 / 2411.6%9 / 5748.7 / Too few decisions245 / 44 decisionsN/A1.781.59 to 2.00 · risks taken 257 / 446 Risk pays · needless 32 / 574 Safety pays+0.70+0.60 to +0.82 · cautious95.3%245 / 257unstable: fewer than 200 needless and missed risks behind a mean cost +0.06β 1.11 · needless risks lost 0.078 on average (32) · missed risks 0.091 (189)
Gemini 4 ArgonhighchatLeague 374.172.0 to 76.150.8%224 / 44197.4%484 / 49788.5%224 / 25345.6%201 / 4412.6%13 / 4973.6%16 / 44112.6needless risks 0.18 · missed risks 8.05 · wrong risks 0.55 · right risks 1.70 · right safety 1.34 · No trade-off 0.75938Risk pays 441 · Safety pays 497 · No trade-off 39466.4%148 / 2230.4%2 / 49711.0 / Too few decisions224 / 29 decisionsN/A2.041.82 to 2.36 · risks taken 240 / 441 Risk pays · needless 13 / 497 Safety pays+0.91+0.77 to +1.09 · cautious93.3%224 / 240unstable: fewer than 200 needless risks behind a mean cost −0.47β 0.38 · needless risks lost 0.036 on average (13) · missed risks 0.107 (201)
GPT-6.1 SollowchatLeague 372.769.9 to 75.551.7%198 / 38393.8%602 / 64278.0%198 / 25444.1%169 / 3836.2%40 / 6424.2%16 / 38312.1needless risks 1.00 · missed risks 6.27 · wrong risks 0.83 · right risks 1.62 · right safety 1.78 · No trade-off 0.621,025Risk pays 383 · Safety pays 642 · No trade-off 31965.0%134 / 2061.6%10 / 642Too few decisions / Too few decisions198 / 56 decisionsN/A1.681.51 to 1.88 · risks taken 214 / 383 Risk pays · needless 40 / 642 Safety pays+0.69+0.58 to +0.83 · cautious92.5%198 / 214unstable: fewer than 200 needless and missed risks behind a mean cost +0.07β 1.13 · needless risks lost 0.067 on average (40) · missed risks 0.100 (169)
Claude Fable 5.1highchatLeague 372.369.8 to 74.750.7%215 / 42494.0%529 / 56377.1%215 / 27942.2%179 / 4246.0%34 / 5637.1%30 / 42414.0needless risks 1.38 · missed risks 6.65 · wrong risks 1.53 · right risks 1.48 · right safety 2.23 · No trade-off 0.68987Risk pays 424 · Safety pays 563 · No trade-off 29965.5%144 / 2202.7%15 / 5639.7 / Too few decisions215 / 64 decisionsN/A1.741.56 to 1.96 · risks taken 245 / 424 Risk pays · needless 34 / 563 Safety pays+0.67+0.55 to +0.81 · cautious87.8%215 / 245unstable: fewer than 200 needless and missed risks behind a mean cost +0.22β 1.46 · needless risks lost 0.105 on average (34) · missed risks 0.096 (179)
GPT-6.1 SolhighchatLeague 372.269.5 to 74.950.6%212 / 41993.7%568 / 60678.2%212 / 27144.4%186 / 4196.3%38 / 6065.0%21 / 4199.9needless risks 0.89 · missed risks 4.75 · wrong risks 0.90 · right risks 1.27 · right safety 1.53 · No trade-off 0.561,025Risk pays 419 · Safety pays 606 · No trade-off 36471.2%148 / 2081.7%10 / 60610.8 / Too few decisions212 / 59 decisionsN/A1.671.42 to 1.94 · risks taken 233 / 419 Risk pays · needless 38 / 606 Safety pays+0.69+0.60 to +0.79 · cautious91.0%212 / 233unstable: fewer than 200 needless and missed risks behind a mean cost +0.17β 1.32 · needless risks lost 0.065 on average (38) · missed risks 0.071 (186)
Claude Opus 5.5highchatLeague 370.067.3 to 72.643.4%199 / 45996.6%590 / 61181.6%199 / 24451.4%236 / 4593.4%21 / 6115.2%24 / 45914.4needless risks 0.62 · missed risks 8.96 · wrong risks 1.23 · right risks 1.26 · right safety 1.55 · No trade-off 0.811,070Risk pays 459 · Safety pays 611 · No trade-off 40756.8%146 / 2571.5%9 / 611Too few decisions / Too few decisions199 / 45 decisionsN/A1.771.53 to 2.05 · risks taken 223 / 459 Risk pays · needless 21 / 611 Safety pays+0.92+0.79 to +1.08 · cautious89.2%199 / 223unstable: fewer than 200 needless risks behind a mean cost +0.02β 1.04 · needless risks lost 0.088 on average (21) · missed risks 0.112 (236)
Gemini 4 ArgonlowchatLeague 369.266.2 to 72.347.8%209 / 43790.7%543 / 59971.6%209 / 29246.0%201 / 4379.3%56 / 5996.2%27 / 43716.5needless risks 2.37 · missed risks 8.03 · wrong risks 1.23 · right risks 1.35 · right safety 2.53 · No trade-off 0.941,036Risk pays 437 · Safety pays 599 · No trade-off 44859.1%137 / 2324.0%24 / 59910.8 / Too few decisions209 / 83 decisionsN/A1.421.22 to 1.64 · risks taken 236 / 437 Risk pays · needless 56 / 599 Safety pays+0.61+0.52 to +0.70 · cautious88.6%209 / 236unstable: fewer than 200 needless risks behind a mean cost +0.26β 1.46 · needless risks lost 0.126 on average (56) · missed risks 0.119 (201)
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first releasechatLeague 269.067.3 to 70.945.0%338 / 75193.0%891 / 95875.1%338 / 45049.0%368 / 7517.0%67 / 9586.0%45 / 75119.0needless risks 2.27 · missed risks 8.56 · wrong risks 1.93 · right risks 2.41 · right safety 2.89 · No trade-off 0.931,709Risk pays 751 · Safety pays 958 · No trade-off 47962.8%255 / 4063.4%33 / 9589.1 / Too few decisions338 / 112 decisions0.53371 risks with a probability‡; 79 split into groups left out; mean confidence on missed risks 100.0% (304)1.501.36 to 1.64 · risks taken 383 / 751 Risk pays · needless 67 / 958 Safety pays+0.72+0.63 to +0.81 · cautious88.3%338 / 383unstable: fewer than 200 needless risks behind a mean cost +0.41β 1.86 · needless risks lost 0.148 on average (67) · missed risks 0.102 (368)
Grok 4.7highchatLeague 368.566.1 to 71.144.0%190 / 43293.0%454 / 48876.6%190 / 24850.5%218 / 4327.0%34 / 4885.6%24 / 43218.8needless risks 1.31 · missed risks 9.66 · wrong risks 2.39 · right risks 3.33 · right safety 1.53 · No trade-off 0.56920Risk pays 432 · Safety pays 488 · No trade-off 41753.8%129 / 2402.7%13 / 488Too few decisions / Too few decisions190 / 58 decisionsN/A1.461.26 to 1.70 · risks taken 214 / 432 Risk pays · needless 34 / 488 Safety pays+0.74+0.64 to +0.86 · cautious88.8%190 / 214unstable: fewer than 200 needless risks behind a mean cost −0.01β 0.99 · needless risks lost 0.103 on average (34) · missed risks 0.118 (218)
DeepSeek-V4.1-Flashlow · API · 1017chatLeague 268.566.4 to 70.642.5%321 / 75594.5%872 / 92377.7%321 / 41352.1%393 / 7555.5%51 / 9235.4%41 / 75517.1needless risks 1.48 · missed risks 9.93 · wrong risks 1.57 · right risks 1.68 · right safety 1.63 · No trade-off 0.861,678Risk pays 755 · Safety pays 923 · No trade-off 54455.4%226 / 4082.8%26 / 9239.2 / Too few decisions321 / 92 decisions0.51340 risks with a probability‡; 73 split into groups left out; mean confidence on missed risks 100.0% (337)1.541.39 to 1.70 · risks taken 362 / 755 Risk pays · needless 51 / 923 Safety pays+0.82+0.74 to +0.91 · cautious88.7%321 / 362unstable: fewer than 200 needless risks behind a mean cost +0.22β 1.40 · needless risks lost 0.129 on average (51) · missed risks 0.112 (393)
Claude Sonnet 5.5highchatLeague 268.266.3 to 70.041.8%308 / 73794.6%841 / 88981.9%308 / 37655.5%409 / 7375.4%48 / 8892.7%20 / 73716.7needless risks 1.12 · missed risks 10.76 · wrong risks 0.64 · right risks 1.70 · right safety 1.58 · No trade-off 0.921,626Risk pays 737 · Safety pays 889 · No trade-off 53856.2%230 / 4091.9%17 / 8898.6 / Too few decisions308 / 68 decisionsN/A1.461.33 to 1.61 · risks taken 328 / 737 Risk pays · needless 48 / 889 Safety pays+0.87+0.80 to +0.95 · cautious93.9%308 / 328unstable: fewer than 200 needless risks behind a mean cost +0.05β 1.07 · needless risks lost 0.101 on average (48) · missed risks 0.114 (409)
Claude Fable 5.1lowchatLeague 267.865.6 to 69.941.0%270 / 65894.6%781 / 82677.4%270 / 34953.8%354 / 6585.4%45 / 8265.2%34 / 65817.9needless risks 1.59 · missed risks 8.82 · wrong risks 1.49 · right risks 1.74 · right safety 2.68 · No trade-off 1.611,484Risk pays 658 · Safety pays 826 · No trade-off 52458.2%198 / 3403.4%28 / 8269.3 / Too few decisions270 / 79 decisionsN/A1.501.33 to 1.70 · risks taken 304 / 658 Risk pays · needless 45 / 826 Safety pays+0.85+0.76 to +0.94 · cautious88.8%270 / 304unstable: fewer than 200 needless risks behind a mean cost +0.38β 1.78 · needless risks lost 0.142 on average (45) · missed risks 0.100 (354)
Grok 4.7lowchatLeague 267.766.0 to 69.542.3%295 / 69793.0%852 / 91671.4%295 / 41349.9%348 / 6977.0%64 / 9167.7%54 / 69719.1needless risks 1.99 · missed risks 10.40 · wrong risks 2.28 · right risks 1.90 · right safety 2.04 · No trade-off 0.511,613Risk pays 697 · Safety pays 916 · No trade-off 53951.4%201 / 3913.4%31 / 9169.3 / Too few decisions295 / 118 decisionsN/A1.481.34 to 1.63 · risks taken 349 / 697 Risk pays · needless 64 / 916 Safety pays+0.74+0.67 to +0.81 · cautious84.5%295 / 349unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.37 · needless risks lost 0.134 on average (64) · missed risks 0.129 (348)
Claude Opus 5.5lowchatLeague 366.964.7 to 69.240.1%187 / 46693.8%541 / 57777.3%187 / 24255.8%260 / 4666.2%36 / 5774.1%19 / 46615.4needless risks 1.67 · missed risks 8.98 · wrong risks 1.10 · right risks 1.00 · right safety 2.14 · No trade-off 0.551,043Risk pays 466 · Safety pays 577 · No trade-off 36853.0%131 / 2473.5%20 / 577Too few decisions / Too few decisions187 / 55 decisionsN/A1.381.17 to 1.62 · risks taken 206 / 466 Risk pays · needless 36 / 577 Safety pays+0.84+0.72 to +0.97 · cautious90.8%187 / 206unstable: fewer than 200 needless risks behind a mean cost +0.37β 1.66 · needless risks lost 0.131 on average (36) · missed risks 0.097 (260)
Gemini 3.8 FlashhighchatLeague 266.965.2 to 68.536.1%264 / 73197.6%858 / 87987.4%264 / 30261.6%450 / 7312.4%21 / 8792.3%17 / 73114.9needless risks 0.42 · missed risks 10.77 · wrong risks 0.59 · right risks 1.16 · right safety 1.50 · No trade-off 0.501,610Risk pays 731 · Safety pays 879 · No trade-off 55852.9%193 / 3650.5%4 / 8799.2 / Too few decisions264 / 38 decisionsN/A1.681.49 to 1.88 · risks taken 281 / 731 Risk pays · needless 21 / 879 Safety pays+1.13+1.05 to +1.23 · cautious94.0%264 / 281unstable: fewer than 200 needless risks behind a mean cost −0.00β 0.99 · needless risks lost 0.086 on average (21) · missed risks 0.104 (450)
Kimi K3.1 ProoffchatLeague 365.963.2 to 68.942.4%153 / 36189.5%416 / 46565.7%153 / 23349.0%177 / 36110.5%49 / 4658.6%31 / 36128.0needless risks 4.62 · missed risks 10.08 · wrong risks 3.16 · right risks 2.22 · right safety 4.08 · No trade-off 3.89826Risk pays 361 · Safety pays 465 · No trade-off 34151.4%109 / 2126.7%31 / 465Too few decisions / Too few decisions153 / 80 decisionsN/A1.271.11 to 1.45 · risks taken 184 / 361 Risk pays · needless 49 / 465 Safety pays+0.61+0.51 to +0.72 · cautious83.2%153 / 184unstable: fewer than 200 needless and missed risks behind a mean cost +0.60β 2.13 · needless risks lost 0.220 on average (49) · missed risks 0.133 (177)
Gemini 3.8 FlashlowchatLeague 265.163.5 to 66.833.8%214 / 63396.5%877 / 90980.8%214 / 26563.2%400 / 6333.5%32 / 9093.0%19 / 63316.3needless risks 0.80 · missed risks 9.90 · wrong risks 0.74 · right risks 0.79 · right safety 2.82 · No trade-off 1.291,542Risk pays 633 · Safety pays 909 · No trade-off 51149.5%151 / 3051.7%15 / 9099.7 / Too few decisions214 / 51 decisionsN/A1.471.33 to 1.64 · risks taken 233 / 633 Risk pays · needless 32 / 909 Safety pays+1.07+0.99 to +1.17 · cautious91.8%214 / 233unstable: fewer than 200 needless risks behind a mean cost +0.25β 1.45 · needless risks lost 0.102 on average (32) · missed risks 0.102 (400)
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 264.562.9 to 66.231.9%223 / 70097.1%878 / 90482.0%223 / 27264.9%454 / 7002.9%26 / 9043.3%23 / 70019.6needless risks 0.54 · missed risks 12.79 · wrong risks 1.21 · right risks 0.96 · right safety 3.00 · No trade-off 1.081,604Risk pays 700 · Safety pays 904 · No trade-off 51843.7%164 / 3751.2%11 / 90410.3 / Too few decisions223 / 49 decisions0.52225 risks with a probability‡; 47 split into groups left out; mean confidence on missed risks 99.7% (385)1.511.36 to 1.69 · risks taken 246 / 700 Risk pays · needless 26 / 904 Safety pays+1.14+1.05 to +1.23 · cautious90.7%223 / 246unstable: fewer than 200 needless risks behind a mean cost −0.03β 0.95 · needless risks lost 0.088 on average (26) · missed risks 0.120 (454)
d1-ultradecisionLeague 264.061.7 to 66.342.0%289 / 68886.0%702 / 81662.7%289 / 46149.6%341 / 68814.0%114 / 8168.4%58 / 68830.3needless risks 3.66 · missed risks 13.25 · wrong risks 3.02 · right risks 2.37 · right safety 4.52 · No trade-off 3.511,504Risk pays 688 · Safety pays 816 · No trade-off 56748.5%180 / 3717.4%60 / 8169.9 / Too few decisions289 / 172 decisions0.64461 risks with a probability‡; mean confidence on missed risks 34.2% (340)1.090.92 to 1.26 · risks taken 347 / 688 Risk pays · needless 114 / 816 Safety pays+0.53+0.46 to +0.61 · cautious83.3%289 / 347unstable: fewer than 200 needless risks behind a mean cost −0.02β 0.98 · needless risks lost 0.133 on average (114) · missed risks 0.161 (341)
Claude Sonnet 5.5lowchatLeague 263.561.6 to 65.430.9%204 / 66196.2%883 / 91877.6%204 / 26365.5%433 / 6613.8%35 / 9183.6%24 / 66120.1needless risks 1.22 · missed risks 12.96 · wrong risks 0.96 · right risks 1.07 · right safety 2.67 · No trade-off 1.261,579Risk pays 661 · Safety pays 918 · No trade-off 57141.9%145 / 3462.7%25 / 9188.7 / Too few decisions204 / 59 decisionsN/A1.371.19 to 1.58 · risks taken 228 / 661 Risk pays · needless 35 / 918 Safety pays+1.08+0.99 to +1.19 · cautious89.5%204 / 228unstable: fewer than 200 needless risks behind a mean cost +0.35β 1.62 · needless risks lost 0.150 on average (35) · missed risks 0.129 (433)
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 262.660.4 to 65.033.2%194 / 58492.1%732 / 79563.4%194 / 30658.4%341 / 5847.9%63 / 7958.4%49 / 58422.5needless risks 2.06 · missed risks 11.12 · wrong risks 2.12 · right risks 1.55 · right safety 3.07 · No trade-off 2.571,379Risk pays 584 · Safety pays 795 · No trade-off 46844.7%136 / 3044.9%39 / 795Too few decisions / Too few decisions194 / 112 decisions0.60261 risks with a probability‡; 45 split into groups left out; mean confidence on missed risks 97.5% (299)1.201.04 to 1.36 · risks taken 243 / 584 Risk pays · needless 63 / 795 Safety pays+0.81+0.72 to +0.90 · cautious79.8%194 / 243unstable: fewer than 200 needless risks behind a mean cost +0.26β 1.36 · needless risks lost 0.121 on average (63) · missed risks 0.120 (341)
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 262.060.1 to 63.933.8%242 / 71590.2%735 / 81564.0%242 / 37858.3%417 / 7159.8%80 / 8157.8%56 / 71529.0needless risks 2.89 · missed risks 14.82 · wrong risks 2.55 · right risks 1.94 · right safety 3.41 · No trade-off 3.431,530Risk pays 715 · Safety pays 815 · No trade-off 57842.2%174 / 4126.0%49 / 81510.0 / Too few decisions242 / 136 decisions0.56332 risks with a probability‡; 46 split into groups left out; mean confidence on missed risks 100.0% (355)1.080.92 to 1.24 · risks taken 298 / 715 Risk pays · needless 80 / 815 Safety pays+0.75+0.66 to +0.84 · cautious81.2%242 / 298unstable: fewer than 200 needless risks behind a mean cost +0.14β 1.16 · needless risks lost 0.153 on average (80) · missed risks 0.150 (417)
GLM-5.3lowchatLeague 261.759.7 to 63.830.7%205 / 66892.7%726 / 78367.7%205 / 30363.2%422 / 6687.3%57 / 7836.1%41 / 66830.9needless risks 1.87 · missed risks 17.40 · wrong risks 2.14 · right risks 1.78 · right safety 3.99 · No trade-off 3.701,451Risk pays 668 · Safety pays 783 · No trade-off 53836.8%141 / 3834.5%35 / 7838.8 / Too few decisions205 / 98 decisions0.52259 risks with a probability‡; 43 split into groups left out; mean confidence on missed risks 91.3% (373)1.120.95 to 1.29 · risks taken 246 / 668 Risk pays · needless 57 / 783 Safety pays+0.89+0.81 to +0.98 · cautious83.3%205 / 246unstable: fewer than 200 needless risks behind a mean cost −0.06β 0.93 · needless risks lost 0.130 on average (57) · missed risks 0.164 (422)
GLiDEAPIdecisionLeague 260.458.3 to 62.636.9%249 / 67584.0%756 / 90052.4%249 / 47551.0%344 / 67516.0%144 / 90012.1%82 / 67533.2needless risks 7.18 · missed risks 10.97 · wrong risks 4.79 · right risks 2.03 · right safety 4.97 · No trade-off 3.241,575Risk pays 675 · Safety pays 900 · No trade-off 58445.7%169 / 37011.9%107 / 90012.2 / 9.9249 / 226 decisions0.58475 risks with a probability‡; mean confidence on missed risks 39.7% (343)0.970.85 to 1.09 · risks taken 331 / 675 Risk pays · needless 144 / 900 Safety pays+0.51+0.43 to +0.58 · cautious75.2%249 / 331unstable: fewer than 200 needless risks behind a mean cost +0.76β 2.08 · needless risks lost 0.215 on average (144) · missed risks 0.138 (344)
Jev 1.13decisionLeague 159.557.4 to 61.627.0%190 / 70592.1%710 / 77161.7%190 / 30865.0%458 / 7057.9%61 / 7718.1%57 / 70534.4needless risks 2.10 · missed risks 21.41 · wrong risks 3.52 · right risks 2.12 · right safety 2.66 · No trade-off 2.601,476Risk pays 705 · Safety pays 771 · No trade-off 43533.0%143 / 4333.8%29 / 771Too few decisions / Too few decisions190 / 118 decisions0.53308 risks with a probability‡; mean confidence on missed risks 52.2% (458)1.020.82 to 1.26 · risks taken 247 / 705 Risk pays · needless 61 / 771 Safety pays+0.90+0.81 to +1.00 · cautious76.9%190 / 247unstable: fewer than 200 needless risks behind a mean cost −0.21β 0.80 · needless risks lost 0.131 on average (61) · missed risks 0.179 (458)
Qwen3.8 FlashoffchatLeague 259.557.2 to 61.734.4%243 / 70684.6%613 / 72554.5%243 / 44652.7%372 / 70615.4%112 / 72512.9%91 / 70633.6needless risks 4.64 · missed risks 14.39 · wrong risks 3.96 · right risks 3.46 · right safety 2.96 · No trade-off 4.171,431Risk pays 706 · Safety pays 725 · No trade-off 59642.5%166 / 3919.5%69 / 7257.7 / 9.5243 / 203 decisions0.57386 risks with a probability‡; 60 split into groups left out; mean confidence on missed risks 54.5% (328)0.950.79 to 1.12 · risks taken 334 / 706 Risk pays · needless 112 / 725 Safety pays+0.54+0.46 to +0.62 · cautious72.8%243 / 334unstable: fewer than 200 needless risks behind a mean cost +0.10β 1.10 · needless risks lost 0.168 on average (112) · missed risks 0.157 (372)
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first releasechatLeague 159.357.1 to 61.529.1%192 / 65989.5%603 / 67461.9%192 / 31063.7%420 / 65910.5%71 / 6747.1%47 / 65935.6needless risks 2.94 · missed risks 20.07 · wrong risks 3.16 · right risks 3.52 · right safety 3.53 · No trade-off 2.341,333Risk pays 659 · Safety pays 674 · No trade-off 40338.3%150 / 3925.6%38 / 674Too few decisions / Too few decisions192 / 118 decisions0.54269 risks with a probability‡; 41 split into groups left out; mean confidence on missed risks 72.5% (378)0.900.74 to 1.06 · risks taken 239 / 659 Risk pays · needless 71 / 674 Safety pays+0.80+0.72 to +0.89 · cautious80.3%192 / 239unstable: fewer than 200 needless risks behind a mean cost −0.13β 0.89 · needless risks lost 0.144 on average (71) · missed risks 0.166 (420)
Jev 1.14decisionLeague 259.357.1 to 61.429.1%184 / 63289.4%657 / 73559.5%184 / 30963.4%401 / 63210.6%78 / 7357.4%47 / 63231.7needless risks 2.84 · missed risks 16.00 · wrong risks 2.91 · right risks 1.72 · right safety 4.37 · No trade-off 3.891,367Risk pays 632 · Safety pays 735 · No trade-off 54533.2%118 / 3556.1%45 / 735Too few decisions / Too few decisions184 / 125 decisions0.49309 risks with a probability‡; mean confidence on missed risks 48.2% (401)0.900.73 to 1.07 · risks taken 231 / 632 Risk pays · needless 78 / 735 Safety pays+0.79+0.73 to +0.86 · cautious79.7%184 / 231unstable: fewer than 200 needless risks behind a mean cost +0.07β 1.06 · needless risks lost 0.139 on average (78) · missed risks 0.153 (401)
GLiDEv1.2 · NVIDIA RTX PRO 6000 BlackwelldecisionLeague 158.856.7 to 60.833.0%219 / 66484.6%698 / 82554.8%219 / 40058.9%391 / 66415.4%127 / 8258.1%54 / 66435.0needless risks 5.15 · missed risks 16.67 · wrong risks 2.72 · right risks 1.56 · right safety 5.81 · No trade-off 3.091,489Risk pays 664 · Safety pays 825 · No trade-off 51438.4%139 / 3629.1%75 / 82512.4 / Too few decisions219 / 181 decisions0.60400 risks with a probability‡; mean confidence on missed risks 35.5% (391)0.790.67 to 0.93 · risks taken 273 / 664 Risk pays · needless 127 / 825 Safety pays+0.62+0.54 to +0.71 · cautious80.2%219 / 273unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.18 · needless risks lost 0.162 on average (127) · missed risks 0.171 (391)
DeepSeek-V4.1-Flashoff · API · 1017chatLeague 258.556.6 to 60.525.4%181 / 71391.7%683 / 74562.8%181 / 28868.3%487 / 7138.3%62 / 7456.3%45 / 71335.2needless risks 2.49 · missed risks 21.43 · wrong risks 2.50 · right risks 1.61 · right safety 3.75 · No trade-off 3.471,458Risk pays 713 · Safety pays 745 · No trade-off 57633.3%138 / 4155.1%38 / 745Too few decisions / Too few decisions181 / 107 decisions0.58262 risks with a probability‡; 26 split into groups left out; mean confidence on missed risks 81.0% (437)0.900.74 to 1.07 · risks taken 226 / 713 Risk pays · needless 62 / 745 Safety pays+0.93+0.84 to +1.02 · cautious80.1%181 / 226unstable: fewer than 200 needless risks behind a mean cost −0.05β 0.96 · needless risks lost 0.164 on average (62) · missed risks 0.179 (487)
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 BlackwelldecisionLeague 157.956.0 to 59.933.4%244 / 73182.5%682 / 82752.4%244 / 46656.1%410 / 73117.5%145 / 82710.5%77 / 73136.9needless risks 6.42 · missed risks 16.25 · wrong risks 3.85 · right risks 2.17 · right safety 4.74 · No trade-off 3.511,558Risk pays 731 · Safety pays 827 · No trade-off 47038.9%162 / 41612.2%101 / 82710.1 / 8.9244 / 222 decisions0.55416 risks with a probability‡; 50 split into groups left out; mean confidence on missed risks 42.4% (337)0.780.67 to 0.89 · risks taken 321 / 731 Risk pays · needless 145 / 827 Safety pays+0.54+0.48 to +0.61 · cautious76.0%244 / 321unstable: fewer than 200 needless risks behind a mean cost +0.30β 1.26 · needless risks lost 0.180 on average (145) · missed risks 0.161 (410)
celeris-1-decisiondecisionLeague 157.855.9 to 59.723.9%163 / 68391.7%749 / 81760.8%163 / 26870.7%483 / 6838.3%68 / 8175.4%37 / 68335.6needless risks 2.30 · missed risks 20.00 · wrong risks 2.07 · right risks 1.37 · right safety 5.50 · No trade-off 4.401,500Risk pays 683 · Safety pays 817 · No trade-off 42232.1%120 / 3744.7%38 / 817Too few decisions / Too few decisions163 / 105 decisions0.52268 risks with a probability‡; mean confidence on missed risks 67.9% (483)0.840.67 to 1.03 · risks taken 200 / 683 Risk pays · needless 68 / 817 Safety pays+0.96+0.88 to +1.05 · cautious81.5%163 / 200unstable: fewer than 200 needless risks behind a mean cost −0.03β 0.97 · needless risks lost 0.130 on average (68) · missed risks 0.159 (483)
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 157.656.1 to 59.023.5%178 / 75691.7%671 / 73263.3%178 / 28170.9%536 / 7568.3%61 / 7325.6%42 / 75645.0needless risks 3.20 · missed risks 26.37 · wrong risks 2.82 · right risks 2.92 · right safety 5.75 · No trade-off 3.981,488Risk pays 756 · Safety pays 732 · No trade-off 51726.3%124 / 4716.3%46 / 732Too few decisions / Too few decisions178 / 103 decisions0.50244 risks with a probability‡; 37 split into groups left out; mean confidence on missed risks 70.5% (477)0.830.70 to 0.98 · risks taken 220 / 756 Risk pays · needless 61 / 732 Safety pays+0.96+0.89 to +1.04 · cautious80.9%178 / 220unstable: fewer than 200 needless risks behind a mean cost +0.04β 1.03 · needless risks lost 0.210 on average (61) · missed risks 0.197 (536)
ClefNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkdecisionLeague 257.555.4 to 59.839.5%263 / 66675.6%603 / 79847.0%263 / 56045.2%301 / 66624.4%195 / 79815.3%102 / 66636.5needless risks 8.85 · missed risks 11.85 · wrong risks 5.48 · right risks 2.26 · right safety 4.18 · No trade-off 3.921,464Risk pays 666 · Safety pays 798 · No trade-off 51447.4%182 / 38415.4%123 / 7989.3 / 10.7263 / 297 decisions0.49552 risks with a probability‡; 8 split into groups left out; mean confidence on missed risks 22.6% (297)0.810.67 to 0.94 · risks taken 365 / 666 Risk pays · needless 195 / 798 Safety pays+0.29+0.20 to +0.37 · cautious72.1%263 / 365unstable: fewer than 200 needless risks behind a mean cost +0.40β 1.38 · needless risks lost 0.180 on average (195) · missed risks 0.156 (301)
CygnetNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkdecisionLeague 157.155.3 to 58.827.3%205 / 75086.8%685 / 78955.9%205 / 36764.9%487 / 75013.2%104 / 7897.7%58 / 75038.6needless risks 4.14 · missed risks 19.98 · wrong risks 3.19 · right risks 2.79 · right safety 4.26 · No trade-off 4.201,539Risk pays 750 · Safety pays 789 · No trade-off 47735.6%159 / 4478.5%67 / 7897.6 / Too few decisions205 / 162 decisions0.56339 risks with a probability‡; 28 split into groups left out; mean confidence on missed risks 65.9% (452)0.730.60 to 0.88 · risks taken 263 / 750 Risk pays · needless 104 / 789 Safety pays+0.75+0.67 to +0.83 · cautious77.9%205 / 263unstable: fewer than 200 needless risks behind a mean cost +0.03β 1.02 · needless risks lost 0.160 on average (104) · missed risks 0.165 (487)
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkdecisionLeague 156.954.9 to 58.833.1%233 / 70380.7%682 / 84550.3%233 / 46357.3%403 / 70319.3%163 / 8459.5%67 / 70338.8needless risks 8.32 · missed risks 16.99 · wrong risks 3.72 · right risks 2.11 · right safety 3.90 · No trade-off 3.751,548Risk pays 703 · Safety pays 845 · No trade-off 50439.9%160 / 40114.2%120 / 84510.1 / 10.1233 / 230 decisions0.55400 risks with a probability‡; 63 split into groups left out; mean confidence on missed risks 53.3% (340)0.680.56 to 0.81 · risks taken 300 / 703 Risk pays · needless 163 / 845 Safety pays+0.53+0.46 to +0.59 · cautious77.7%233 / 300unstable: fewer than 200 needless risks behind a mean cost +0.55β 1.45 · needless risks lost 0.209 on average (163) · missed risks 0.173 (403)
SubQ Snap PreviewdecisionLeague 156.553.8 to 59.137.9%297 / 78375.1%673 / 89645.9%297 / 64745.8%359 / 78324.9%223 / 89616.2%127 / 78343.9needless risks 10.40 · missed risks 14.35 · wrong risks 7.22 · right risks 3.21 · right safety 5.11 · No trade-off 3.581,679Risk pays 783 · Safety pays 896 · No trade-off 52145.9%201 / 43817.4%156 / 8969.6 / 10.9297 / 350 decisions0.61647 risks with a probability‡; mean confidence on missed risks 44.3% (358)0.780.65 to 0.93 · risks taken 424 / 783 Risk pays · needless 223 / 896 Safety pays+0.29+0.22 to +0.36 · cautious70.0%297 / 424+0.37β 1.33 · needless risks lost 0.205 on average (223) · missed risks 0.176 (359)
GPT-6 LunaDecisions APIdecisionLeague 155.753.8 to 57.621.6%160 / 74089.7%636 / 70956.3%160 / 28471.5%529 / 74010.3%73 / 7096.9%51 / 74043.3needless risks 3.67 · missed risks 25.05 · wrong risks 3.45 · right risks 2.58 · right safety 4.50 · No trade-off 4.071,449Risk pays 740 · Safety pays 709 · No trade-off 51328.6%123 / 4307.5%53 / 709Too few decisions / Too few decisions160 / 124 decisions0.60284 risks with a probability‡; mean confidence on missed risks 54.6% (529)0.690.54 to 0.87 · risks taken 211 / 740 Risk pays · needless 73 / 709 Safety pays+0.91+0.83 to +1.00 · cautious75.8%160 / 211unstable: fewer than 200 needless risks behind a mean cost +0.02β 1.02 · needless risks lost 0.197 on average (73) · missed risks 0.186 (529)
Winnow-12BNVIDIA RTX PRO 6000 BlackwelldecisionLeague 155.252.9 to 57.233.7%263 / 78076.6%573 / 74848.7%263 / 54053.2%415 / 78023.4%175 / 74813.1%102 / 78049.7needless risks 10.05 · missed risks 19.93 · wrong risks 6.06 · right risks 3.45 · right safety 5.37 · No trade-off 4.861,528Risk pays 780 · Safety pays 748 · No trade-off 48739.9%176 / 44116.4%123 / 7489.8 / 10.2263 / 277 decisions0.62518 risks with a probability‡; 22 split into groups left out; mean confidence on missed risks 59.7% (390)0.640.53 to 0.77 · risks taken 365 / 780 Risk pays · needless 175 / 748 Safety pays+0.40+0.32 to +0.49 · cautious72.1%263 / 365unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.15 · needless risks lost 0.231 on average (175) · missed risks 0.194 (415)
Mercury DecidedecisionLeague 153.451.5 to 55.218.7%129 / 69088.1%583 / 66251.2%129 / 25274.9%517 / 69011.9%79 / 6626.4%44 / 69047.9needless risks 5.59 · missed risks 29.17 · wrong risks 2.52 · right risks 1.67 · right safety 4.44 · No trade-off 4.551,352Risk pays 690 · Safety pays 662 · No trade-off 50121.8%91 / 4177.7%51 / 662Too few decisions / Too few decisions129 / 123 decisions0.63252 risks with a probability‡; mean confidence on missed risks 83.3% (517)0.500.36 to 0.67 · risks taken 173 / 690 Risk pays · needless 79 / 662 Safety pays+0.92+0.84 to +1.02 · cautious74.6%129 / 173unstable: fewer than 200 needless risks behind a mean cost +0.37β 1.20 · needless risks lost 0.262 on average (79) · missed risks 0.209 (517)
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 152.450.3 to 54.531.5%248 / 78873.4%647 / 88238.2%248 / 65047.3%373 / 78826.6%235 / 88221.2%167 / 78847.7needless risks 13.29 · missed risks 12.21 · wrong risks 10.08 · right risks 2.73 · right safety 4.23 · No trade-off 5.151,670Risk pays 788 · Safety pays 882 · No trade-off 52344.8%182 / 40619.2%169 / 88210.5 / 12.3248 / 402 decisions0.52572 risks with a probability‡; 78 split into groups left out; mean confidence on missed risks 100.0% (309)0.690.56 to 0.82 · risks taken 415 / 788 Risk pays · needless 235 / 882 Safety pays+0.28+0.21 to +0.34 · cautious59.8%248 / 415+0.96β 1.93 · needless risks lost 0.248 on average (235) · missed risks 0.144 (373)
Perplexity Decider Flash v1.2NVIDIA RTX PRO 6000 BlackwelldecisionLeague 149.146.5 to 51.631.3%223 / 71266.9%519 / 77633.5%223 / 66542.7%304 / 71233.1%257 / 77626.0%185 / 71259.6needless risks 16.04 · missed risks 13.43 · wrong risks 13.80 · right risks 3.10 · right safety 6.17 · No trade-off 7.111,488Risk pays 712 · Safety pays 776 · No trade-off 51841.9%167 / 39925.3%196 / 77610.7 / 13.4223 / 442 decisions0.56565 risks with a probability‡; 100 split into groups left out; mean confidence on missed risks 32.5% (270)0.620.51 to 0.74 · risks taken 408 / 712 Risk pays · needless 257 / 776 Safety pays+0.13+0.05 to +0.20 · cautious54.7%223 / 408+0.70β 1.54 · needless risks lost 0.250 on average (257) · missed risks 0.177 (304)
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 BlackwelldecisionLeague 148.546.1 to 50.937.3%264 / 70859.8%455 / 76134.4%264 / 76834.7%246 / 70840.2%306 / 76128.0%198 / 70865.6needless risks 21.42 · missed risks 12.49 · wrong risks 17.29 · right risks 3.05 · right safety 4.95 · No trade-off 6.351,469Risk pays 708 · Safety pays 761 · No trade-off 53342.8%170 / 39731.0%236 / 76111.2 / 11.9264 / 504 decisions0.53768 risks with a probability‡; mean confidence on missed risks 14.9% (246)0.640.52 to 0.76 · risks taken 462 / 708 Risk pays · needless 306 / 761 Safety pays−0.07−0.14 to −0.01 · bold57.1%264 / 462+0.62β 1.48 · needless risks lost 0.280 on average (306) · missed risks 0.203 (246)
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 BlackwellchatLeague 145.743.2 to 48.128.8%206 / 71562.7%470 / 75031.8%206 / 64848.5%347 / 71537.3%280 / 75022.7%162 / 71561.5needless risks 18.70 · missed risks 14.79 · wrong risks 13.15 · right risks 2.81 · right safety 5.11 · No trade-off 6.901,465Risk pays 715 · Safety pays 750 · No trade-off 48735.6%147 / 41327.5%206 / 7509.9 / 12.8206 / 442 decisions0.52573 risks with a probability‡; 75 split into groups left out; mean confidence on missed risks 39.2% (316)0.360.23 to 0.49 · risks taken 368 / 715 Risk pays · needless 280 / 750 Safety pays+0.14+0.06 to +0.22 · cautious56.0%206 / 368+1.38β 1.64 · needless risks lost 0.261 on average (280) · missed risks 0.166 (347)
GNUbg weakprogram · reference row, never ranked · x86-64 build: its noise is shifted, so it overplays gammonsLeague 167.163.8 to 70.366.6%410 / 61667.6%533 / 78850.9%410 / 80510.7%66 / 61632.4%255 / 78822.7%140 / 61621.4needless risks 8.80 · missed risks 1.22 · wrong risks 5.99 · right risks 2.66 · right safety 1.66 · No trade-off 1.071,404Risk pays 616 · Safety pays 788 · No trade-off 34082.7%254 / 30718.0%142 / 78811.2 / 13.1410 / 395 decisionsN/A1.701.53 to 1.89 · risks taken 550 / 616 Risk pays · needless 255 / 788 Safety pays−0.39−0.48 to −0.30 · bold74.5%410 / 550unstable: fewer than 200 missed risks behind a mean cost +0.51β 2.39 · needless risks lost 0.120 on average (255) · missed risks 0.065 (66)
Randomprogram · reference row, never ranked · a random legal playLeague 124.922.3 to 27.424.8%112 / 45225.0%121 / 48415.4%112 / 72819.2%87 / 45275.0%363 / 48456.0%253 / 452121.1needless risks 54.22 · missed risks 8.20 · wrong risks 38.00 · right risks 3.36 · right safety 3.88 · No trade-off 13.44936Risk pays 452 · Safety pays 484 · No trade-off 40129.2%78 / 26768.0%329 / 484Too few decisions / 16.4112 / 616 decisionsN/A0.190.05 to 0.34 · risks taken 365 / 452 Risk pays · needless 363 / 484 Safety pays−0.77−0.89 to −0.65 · bold30.7%112 / 365unstable: fewer than 200 missed risks behind a mean cost +2.73β 1.70 · needless risks lost 0.399 on average (363) · missed risks 0.252 (87)
Safestprogram · reference row, never ranked · always plays safeLeague 150.050.0 to 50.00.0%0 / 785100.0%749 / 749N/A0 / 0100.0%785 / 7850.0%0 / 7490.0%0 / 78551.6needless risks 0.00 · missed risks 34.95 · wrong risks 0.00 · right risks 0.00 · right safety 8.45 · No trade-off 8.251,534Risk pays 785 · Safety pays 749 · No trade-off 5240.0%0 / 4600.0%0 / 7490 / 00 / 0 decisionsN/A−0.01−0.05 to 0.02 · risks taken 0 / 785 Risk pays · needless 0 / 749 Safety pays+3.22+3.20 to +3.23 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.183 (785)
GNUbg weakprogram · reference row, never ranked · x86-64 build: its noise is shifted, so it overplays gammonsLeague 262.359.5 to 64.862.7%483 / 77061.9%558 / 90248.2%483 / 1,00314.4%111 / 77038.1%344 / 90222.9%176 / 77027.0needless risks 12.05 · missed risks 2.19 · wrong risks 6.03 · right risks 2.93 · right safety 1.97 · No trade-off 1.801,672Risk pays 770 · Safety pays 902 · No trade-off 46482.4%327 / 39723.4%211 / 90211.2 / 13.5483 / 520 decisionsN/A1.361.22 to 1.50 · risks taken 659 / 770 Risk pays · needless 344 / 902 Safety pays−0.38−0.46 to −0.30 · bold73.3%483 / 659unstable: fewer than 200 missed risks behind a mean cost +0.54β 2.08 · needless risks lost 0.150 on average (344) · missed risks 0.084 (111)
Random (computed)program · reference row, never ranked · computed: every legal play equally likelyLeague 227.026.7 to 27.225.1%3,300.3 / 13,14028.8%4,655.0 / 16,13614.7%3,300.3 / 22,427.916.7%2,193.1 / 13,14071.2%11,481.0 / 16,13658.2%7,646.6 / 13,140117.9needless risks 57.31 · missed risks 5.40 · wrong risks 40.03 · right risks 3.23 · right safety 3.91 · No trade-off 8.0429,276Risk pays 13,140 · Safety pays 16,136 · No trade-off 10,21231.6%2,248.9 / 7,12763.1%10,184.8 / 16,13615.4 / 17.83,300.3 / 19,127.7 decisionsN/A0.410.39 to 0.42 · risks taken 10,946.9 / 13,140 Risk pays · needless 11,481.0 / 16,136 Safety pays−0.76−0.77 to −0.75 · bold30.1%3,300.3 / 10,946.9+2.23β 2.49 · needless risks lost 0.394 on average (11,481.0) · missed risks 0.195 (2,193.1)
Safest (computed)program · reference row, never ranked · computed: always the best safe playLeague 250.050.0 to 50.00.0%0 / 13,140100.0%16,136 / 16,136N/A0 / 0100.0%13,140 / 13,1400.0%0 / 16,1360.0%0 / 13,14025.2needless risks 0.00 · missed risks 25.23 · wrong risks 0.00 · right risks 0.00 · right safety 0.00 · No trade-off 0.0029,276Risk pays 13,140 · Safety pays 16,136 · No trade-off 10,2120.0%0 / 7,1270.0%0 / 16,1360 / 00 / 0 decisionsN/A0.050.04 to 0.06 · risks taken 0 / 13,140 Risk pays · needless 0 / 16,136 Safety pays+3.98+3.98 to +3.98 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.152 (13,140)
GNUbg weakprogram · reference row, never ranked · x86-64 build: its noise is shifted, so it overplays gammonsLeague 362.158.2 to 65.958.8%288 / 49065.5%415 / 63445.9%288 / 62716.7%82 / 49034.5%219 / 63424.5%120 / 49028.1needless risks 11.03 · missed risks 2.16 · wrong risks 6.12 · right risks 4.03 · right safety 2.52 · No trade-off 2.221,124Risk pays 490 · Safety pays 634 · No trade-off 35582.8%202 / 24421.5%136 / 63412.5 / 14.5288 / 339 decisionsN/A1.361.16 to 1.57 · risks taken 408 / 490 Risk pays · needless 219 / 634 Safety pays−0.28−0.36 to −0.20 · bold70.6%288 / 408unstable: fewer than 200 missed risks behind a mean cost +0.67β 2.47 · needless risks lost 0.149 on average (219) · missed risks 0.078 (82)
Random (computed)program · reference row, never ranked · computed: every legal play equally likelyLeague 327.627.1 to 28.025.1%1,293.3 / 5,15330.1%2,036.5 / 6,77114.4%1,293.3 / 8,984.917.5%902.5 / 5,15369.9%4,734.5 / 6,77157.4%2,957.2 / 5,153117.6needless risks 59.90 · missed risks 5.67 · wrong risks 36.45 · right risks 3.08 · right safety 4.60 · No trade-off 7.9411,924Risk pays 5,153 · Safety pays 6,771 · No trade-off 4,39531.4%862.9 / 2,75262.8%4,253.1 / 6,77115.1 / 17.71,293.3 / 7,691.6 decisionsN/A0.410.39 to 0.44 · risks taken 4,250.5 / 5,153 Risk pays · needless 4,734.5 / 6,771 Safety pays−0.73−0.74 to −0.71 · bold30.4%1,293.3 / 4,250.5+2.36β 2.65 · needless risks lost 0.413 on average (4,734.5) · missed risks 0.205 (902.5)
Safest (computed)program · reference row, never ranked · computed: always the best safe playLeague 350.050.0 to 50.00.0%0 / 5,153100.0%6,771 / 6,771N/A0 / 0100.0%5,153 / 5,1530.0%0 / 6,7710.0%0 / 5,15323.5needless risks 0.00 · missed risks 23.47 · wrong risks 0.00 · right risks 0.00 · right safety 0.00 · No trade-off 0.0011,924Risk pays 5,153 · Safety pays 6,771 · No trade-off 4,3950.0%0 / 2,7520.0%0 / 6,7710 / 00 / 0 decisionsN/A0.070.06 to 0.08 · risks taken 0 / 5,153 Risk pays · needless 0 / 6,771 Safety pays+3.76+3.75 to +3.77 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.149 (5,153)

Hit risk only: being hit. Gammon risk and priming are not in it.

Setups met different positions; risk judgement is not adjusted for how hard they were. Paired comparisons wait for the shared positions.

Big risks found: stake above 0.08. Big needless risks: the chosen play lost more than 0.08.

‡ Risks with a probability: decisions whose plays were split into groups have no single probability and are left out; their count is shown per row. Setups that return no probabilities show N/A. Confidence on the call: ranking score only, no calibration gap (the probability is for the chosen play, not the risk). Definitions

Ability, caution and the SDT reference: A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity. Rates corrected by adding 0.5 to the counts and 1 to the totals; a setup without Risk pays or Safety pays decisions shows N/A. Intervals: 95%, blocks of two seat-swapped games resampled. Definitions

The same at the 0.02 margin (robustness)

Swipe for more columns →

All leagues: every 0.001 in the definitions replaced by 0.02
Model and settingLeagueRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
Gemini 4 ArgonhighchatLeague 375.973.4 to 78.156.4%215 / 38195.3%531 / 55785.0%215 / 25340.4%154 / 3814.7%26 / 5573.1%12 / 3811.911.73 to 2.11 · risks taken 227 / 381 Risk pays · needless 26 / 557 Safety pays+0.71+0.59 to +0.84 · cautious94.7%215 / 227unstable: fewer than 200 needless and missed risks behind a mean cost −0.49β 0.40 · needless risks lost 0.036 on average (26) · missed risks 0.132 (154)
GPT-6 AstralowchatLeague 374.971.9 to 77.958.1%209 / 36091.8%505 / 55077.7%209 / 26937.8%136 / 3608.2%45 / 5504.2%15 / 3601.701.52 to 1.88 · risks taken 224 / 360 Risk pays · needless 45 / 550 Safety pays+0.54+0.47 to +0.62 · cautious93.3%209 / 224unstable: fewer than 200 needless and missed risks behind a mean cost −0.44β 0.47 · needless risks lost 0.037 on average (45) · missed risks 0.118 (136)
GPT-6 AstramediumchatLeague 374.972.4 to 77.356.5%225 / 39893.2%580 / 62277.9%225 / 28937.9%151 / 3986.8%42 / 6225.5%22 / 3981.801.62 to 1.99 · risks taken 247 / 398 Risk pays · needless 42 / 622 Safety pays+0.59+0.50 to +0.70 · cautious91.1%225 / 247unstable: fewer than 200 needless and missed risks behind a mean cost −0.09β 0.85 · needless risks lost 0.059 on average (42) · missed risks 0.109 (151)
GPT-6.1 SollowchatLeague 374.270.9 to 77.456.2%182 / 32492.2%646 / 70171.7%182 / 25438.6%125 / 3247.8%55 / 7015.2%17 / 3241.701.49 to 1.90 · risks taken 199 / 324 Risk pays · needless 55 / 701 Safety pays+0.56+0.46 to +0.68 · cautious91.5%182 / 199unstable: fewer than 200 needless and missed risks behind a mean cost −0.06β 0.90 · needless risks lost 0.054 on average (55) · missed risks 0.129 (125)
GPT-6.1 SolhighchatLeague 372.469.7 to 75.152.4%188 / 35992.5%616 / 66669.4%188 / 27138.4%138 / 3597.5%50 / 6669.2%33 / 3591.731.52 to 1.96 · risks taken 221 / 359 Risk pays · needless 50 / 666 Safety pays+0.57+0.48 to +0.67 · cautious85.1%188 / 221unstable: fewer than 200 needless and missed risks behind a mean cost +0.07β 1.12 · needless risks lost 0.054 on average (50) · missed risks 0.090 (138)
Claude Fable 5.1highchatLeague 372.169.5 to 74.952.7%195 / 37091.6%565 / 61769.9%195 / 27938.6%143 / 3708.4%52 / 6178.6%32 / 3701.661.49 to 1.86 · risks taken 227 / 370 Risk pays · needless 52 / 617 Safety pays+0.54+0.43 to +0.67 · cautious85.9%195 / 227unstable: fewer than 200 needless and missed risks behind a mean cost +0.06β 1.11 · needless risks lost 0.076 on average (52) · missed risks 0.114 (143)
Claude Opus 5.5highchatLeague 370.867.8 to 73.846.3%188 / 40695.3%633 / 66477.0%188 / 24447.5%193 / 4064.7%31 / 6646.2%25 / 4061.731.49 to 2.01 · risks taken 213 / 406 Risk pays · needless 31 / 664 Safety pays+0.80+0.68 to +0.94 · cautious88.3%188 / 213unstable: fewer than 200 needless and missed risks behind a mean cost −0.14β 0.78 · needless risks lost 0.064 on average (31) · missed risks 0.134 (193)
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first releasechatLeague 270.268.2 to 72.249.2%315 / 64091.2%975 / 1,06970.0%315 / 45044.4%284 / 6408.8%94 / 1,0696.4%41 / 6401.491.36 to 1.63 · risks taken 356 / 640 Risk pays · needless 94 / 1,069 Safety pays+0.60+0.53 to +0.68 · cautious88.5%315 / 356unstable: fewer than 200 needless risks behind a mean cost +0.33β 1.64 · needless risks lost 0.125 on average (94) · missed risks 0.127 (284)
Claude Sonnet 5.5highchatLeague 269.067.0 to 71.144.0%292 / 66494.0%904 / 96277.7%292 / 37652.1%346 / 6646.0%58 / 9623.9%26 / 6641.501.37 to 1.65 · risks taken 318 / 664 Risk pays · needless 58 / 962 Safety pays+0.80+0.72 to +0.89 · cautious91.8%292 / 318unstable: fewer than 200 needless risks behind a mean cost −0.05β 0.93 · needless risks lost 0.084 on average (58) · missed risks 0.132 (346)
DeepSeek-V4.1-Flashlow · API · 1017chatLeague 268.966.5 to 71.344.8%293 / 65493.0%952 / 1,02470.9%293 / 41347.9%313 / 6547.0%72 / 1,0247.3%48 / 6541.521.36 to 1.70 · risks taken 341 / 654 Risk pays · needless 72 / 1,024 Safety pays+0.71+0.63 to +0.79 · cautious85.9%293 / 341unstable: fewer than 200 needless risks behind a mean cost +0.08β 1.13 · needless risks lost 0.097 on average (72) · missed risks 0.135 (313)
Gemini 4 ArgonlowchatLeague 368.765.6 to 71.949.6%188 / 37987.8%577 / 65764.4%188 / 29244.1%167 / 37912.2%80 / 6576.3%24 / 3791.311.13 to 1.50 · risks taken 212 / 379 Risk pays · needless 80 / 657 Safety pays+0.51+0.42 to +0.60 · cautious88.7%188 / 212unstable: fewer than 200 needless and missed risks behind a mean cost +0.10β 1.14 · needless risks lost 0.092 on average (80) · missed risks 0.140 (167)
Claude Fable 5.1lowchatLeague 268.566.3 to 70.843.7%248 / 56793.2%855 / 91771.1%248 / 34949.4%280 / 5676.8%62 / 9176.9%39 / 5671.511.35 to 1.68 · risks taken 287 / 567 Risk pays · needless 62 / 917 Safety pays+0.74+0.65 to +0.83 · cautious86.4%248 / 287unstable: fewer than 200 needless risks behind a mean cost +0.33β 1.65 · needless risks lost 0.122 on average (62) · missed risks 0.120 (280)
Grok 4.7highchatLeague 367.964.9 to 70.944.6%171 / 38391.1%489 / 53769.0%171 / 24847.8%183 / 3838.9%48 / 5377.6%29 / 3831.401.20 to 1.60 · risks taken 200 / 383 Risk pays · needless 48 / 537 Safety pays+0.64+0.55 to +0.75 · cautious85.5%171 / 200unstable: fewer than 200 needless and missed risks behind a mean cost −0.13β 0.83 · needless risks lost 0.082 on average (48) · missed risks 0.137 (183)
Gemini 3.8 FlashhighchatLeague 267.665.5 to 69.639.2%239 / 61096.1%961 / 1,00079.1%239 / 30256.9%347 / 6103.9%39 / 1,0003.9%24 / 6101.581.40 to 1.79 · risks taken 263 / 610 Risk pays · needless 39 / 1,000 Safety pays+0.97+0.89 to +1.05 · cautious90.9%239 / 263unstable: fewer than 200 needless risks behind a mean cost −0.29β 0.63 · needless risks lost 0.049 on average (39) · missed risks 0.128 (347)
Claude Opus 5.5lowchatLeague 367.164.5 to 69.541.9%168 / 40192.4%593 / 64269.4%168 / 24251.9%208 / 4017.6%49 / 6426.2%25 / 4011.381.17 to 1.61 · risks taken 193 / 401 Risk pays · needless 49 / 642 Safety pays+0.74+0.63 to +0.87 · cautious87.0%168 / 193unstable: fewer than 200 needless risks behind a mean cost +0.30β 1.51 · needless risks lost 0.108 on average (49) · missed risks 0.115 (208)
Grok 4.7lowchatLeague 267.064.9 to 69.142.8%263 / 61491.1%910 / 99963.7%263 / 41347.2%290 / 6148.9%89 / 9999.9%61 / 6141.411.26 to 1.57 · risks taken 324 / 614 Risk pays · needless 89 / 999 Safety pays+0.64+0.57 to +0.71 · cautious81.2%263 / 324unstable: fewer than 200 needless risks behind a mean cost +0.08β 1.13 · needless risks lost 0.104 on average (89) · missed risks 0.150 (290)
Gemini 3.8 FlashlowchatLeague 266.864.9 to 68.837.6%203 / 54096.0%962 / 1,00276.6%203 / 26558.3%315 / 5404.0%40 / 1,0024.1%22 / 5401.541.38 to 1.71 · risks taken 225 / 540 Risk pays · needless 40 / 1,002 Safety pays+0.98+0.90 to +1.07 · cautious90.2%203 / 225unstable: fewer than 200 needless risks behind a mean cost +0.22β 1.40 · needless risks lost 0.093 on average (40) · missed risks 0.124 (315)
Kimi K3.1 ProoffchatLeague 366.763.7 to 69.844.7%143 / 32088.7%449 / 50661.4%143 / 23345.0%144 / 32011.3%57 / 50610.3%33 / 3201.331.18 to 1.49 · risks taken 176 / 320 Risk pays · needless 57 / 506 Safety pays+0.54+0.44 to +0.64 · cautious81.2%143 / 176unstable: fewer than 200 needless and missed risks behind a mean cost +0.55β 2.09 · needless risks lost 0.209 on average (57) · missed risks 0.158 (144)
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 265.663.8 to 67.534.7%213 / 61396.5%956 / 99178.3%213 / 27261.3%376 / 6133.5%35 / 9913.9%24 / 6131.511.35 to 1.68 · risks taken 237 / 613 Risk pays · needless 35 / 991 Safety pays+1.04+0.96 to +1.14 · cautious89.9%213 / 237unstable: fewer than 200 needless risks behind a mean cost +0.07β 1.11 · needless risks lost 0.095 on average (35) · missed risks 0.139 (376)
Claude Sonnet 5.5lowchatLeague 264.262.1 to 66.233.1%190 / 57495.2%957 / 1,00572.2%190 / 26362.5%359 / 5744.8%48 / 1,0054.4%25 / 5741.341.18 to 1.53 · risks taken 215 / 574 Risk pays · needless 48 / 1,005 Safety pays+0.99+0.91 to +1.09 · cautious88.4%190 / 215unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.33 · needless risks lost 0.115 on average (48) · missed risks 0.151 (359)
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 263.261.0 to 65.635.5%182 / 51390.9%787 / 86659.5%182 / 30655.8%286 / 5139.1%79 / 8668.8%45 / 5131.191.03 to 1.35 · risks taken 227 / 513 Risk pays · needless 79 / 866 Safety pays+0.74+0.66 to +0.82 · cautious80.2%182 / 227unstable: fewer than 200 needless risks behind a mean cost +0.22β 1.30 · needless risks lost 0.106 on average (79) · missed risks 0.138 (286)
d1-ultradecisionLeague 263.060.8 to 65.341.8%254 / 60784.2%755 / 89755.1%254 / 46147.4%288 / 60715.8%142 / 89710.7%65 / 6071.060.91 to 1.22 · risks taken 319 / 607 Risk pays · needless 142 / 897 Safety pays+0.47+0.40 to +0.54 · cautious79.6%254 / 319unstable: fewer than 200 needless risks behind a mean cost −0.02β 0.98 · needless risks lost 0.122 on average (142) · missed risks 0.185 (288)
GLM-5.3lowchatLeague 261.659.5 to 63.931.8%184 / 57991.4%797 / 87260.7%184 / 30360.6%351 / 5798.6%75 / 8727.6%44 / 5791.090.91 to 1.27 · risks taken 228 / 579 Risk pays · needless 75 / 872 Safety pays+0.82+0.74 to +0.89 · cautious80.7%184 / 228unstable: fewer than 200 needless risks behind a mean cost −0.12β 0.88 · needless risks lost 0.113 on average (75) · missed risks 0.193 (351)
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 261.259.2 to 63.234.4%217 / 63087.9%791 / 90057.4%217 / 37857.3%361 / 63012.1%109 / 9008.3%52 / 6300.980.83 to 1.15 · risks taken 269 / 630 Risk pays · needless 109 / 900 Safety pays+0.68+0.59 to +0.77 · cautious80.7%217 / 269unstable: fewer than 200 needless risks behind a mean cost +0.11β 1.11 · needless risks lost 0.131 on average (109) · missed risks 0.169 (361)
GLiDEAPIdecisionLeague 260.558.2 to 62.638.4%228 / 59382.5%810 / 98248.0%228 / 47548.9%290 / 59317.5%172 / 98212.6%75 / 5930.960.86 to 1.07 · risks taken 303 / 593 Risk pays · needless 172 / 982 Safety pays+0.45+0.38 to +0.53 · cautious75.2%228 / 303unstable: fewer than 200 needless risks behind a mean cost +0.73β 2.01 · needless risks lost 0.190 on average (172) · missed risks 0.157 (290)
Qwen3.8 FlashoffchatLeague 259.357.1 to 61.535.5%222 / 62583.0%669 / 80649.8%222 / 44650.6%316 / 62517.0%137 / 80613.9%87 / 6250.940.79 to 1.08 · risks taken 309 / 625 Risk pays · needless 137 / 806 Safety pays+0.48+0.41 to +0.56 · cautious71.8%222 / 309unstable: fewer than 200 needless risks behind a mean cost +0.06β 1.05 · needless risks lost 0.147 on average (137) · missed risks 0.180 (316)
Jev 1.14decisionLeague 259.257.0 to 61.429.9%167 / 55988.5%715 / 80854.0%167 / 30961.4%343 / 55911.5%93 / 8088.8%49 / 5590.910.75 to 1.08 · risks taken 216 / 559 Risk pays · needless 93 / 808 Safety pays+0.74+0.68 to +0.81 · cautious77.3%167 / 216unstable: fewer than 200 needless risks behind a mean cost −0.00β 1.00 · needless risks lost 0.121 on average (93) · missed risks 0.174 (343)
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first releasechatLeague 159.056.7 to 61.330.1%177 / 58887.9%655 / 74557.1%177 / 31062.6%368 / 58812.1%90 / 7457.3%43 / 5880.850.72 to 0.98 · risks taken 220 / 588 Risk pays · needless 90 / 745 Safety pays+0.74+0.66 to +0.83 · cautious80.5%177 / 220unstable: fewer than 200 needless risks behind a mean cost −0.20β 0.84 · needless risks lost 0.124 on average (90) · missed risks 0.187 (368)
DeepSeek-V4.1-Flashoff · API · 1017chatLeague 258.856.7 to 60.926.5%168 / 63591.1%750 / 82358.3%168 / 28866.1%420 / 6358.9%73 / 8237.4%47 / 6350.930.77 to 1.10 · risks taken 215 / 635 Risk pays · needless 73 / 823 Safety pays+0.88+0.79 to +0.97 · cautious78.1%168 / 215unstable: fewer than 200 needless risks behind a mean cost −0.09β 0.92 · needless risks lost 0.145 on average (73) · missed risks 0.204 (420)
Jev 1.13decisionLeague 158.456.2 to 60.826.8%167 / 62290.0%769 / 85454.2%167 / 30864.1%399 / 62210.0%85 / 8549.0%56 / 6220.920.72 to 1.13 · risks taken 223 / 622 Risk pays · needless 85 / 854 Safety pays+0.82+0.74 to +0.91 · cautious74.9%167 / 223unstable: fewer than 200 needless risks behind a mean cost −0.36β 0.72 · needless risks lost 0.106 on average (85) · missed risks 0.202 (399)
GLiDEv1.2 · NVIDIA RTX PRO 6000 BlackwelldecisionLeague 157.955.4 to 60.433.1%192 / 58082.7%752 / 90948.0%192 / 40058.1%337 / 58017.3%157 / 9098.8%51 / 5800.740.59 to 0.89 · risks taken 243 / 580 Risk pays · needless 157 / 909 Safety pays+0.57+0.49 to +0.66 · cautious79.0%192 / 243unstable: fewer than 200 needless risks behind a mean cost +0.15β 1.11 · needless risks lost 0.138 on average (157) · missed risks 0.194 (337)
ClefNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkdecisionLeague 257.655.2 to 60.040.6%240 / 59174.6%651 / 87342.9%240 / 56042.8%253 / 59125.4%222 / 87316.6%98 / 5910.840.70 to 0.99 · risks taken 338 / 591 Risk pays · needless 222 / 873 Safety pays+0.24+0.15 to +0.32 · cautious71.0%240 / 338+0.39β 1.39 · needless risks lost 0.168 on average (222) · missed risks 0.178 (253)
CygnetNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkdecisionLeague 157.355.4 to 59.228.2%192 / 68086.4%742 / 85952.3%192 / 36763.2%430 / 68013.6%117 / 8598.5%58 / 6800.760.63 to 0.89 · risks taken 250 / 680 Risk pays · needless 117 / 859 Safety pays+0.72+0.64 to +0.80 · cautious76.8%192 / 250unstable: fewer than 200 needless risks behind a mean cost +0.05β 1.04 · needless risks lost 0.151 on average (117) · missed risks 0.184 (430)
celeris-1-decisiondecisionLeague 157.355.4 to 59.124.1%141 / 58690.5%827 / 91452.6%141 / 26869.1%405 / 5869.5%87 / 9146.8%40 / 5860.810.65 to 0.98 · risks taken 181 / 586 Risk pays · needless 87 / 914 Safety pays+0.90+0.82 to +0.99 · cautious77.9%141 / 181unstable: fewer than 200 needless risks behind a mean cost −0.10β 0.92 · needless risks lost 0.107 on average (87) · missed risks 0.181 (405)
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 156.755.1 to 58.222.8%157 / 68890.6%725 / 80055.9%157 / 28170.1%482 / 6889.4%75 / 8007.1%49 / 6880.790.66 to 0.93 · risks taken 206 / 688 Risk pays · needless 75 / 800 Safety pays+0.92+0.85 to +0.99 · cautious76.2%157 / 206unstable: fewer than 200 needless risks behind a mean cost −0.03β 0.98 · needless risks lost 0.180 on average (75) · missed risks 0.215 (482)
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkdecisionLeague 156.754.7 to 58.733.7%211 / 62679.7%735 / 92245.6%211 / 46355.9%350 / 62620.3%187 / 92210.4%65 / 6260.680.55 to 0.82 · risks taken 276 / 626 Risk pays · needless 187 / 922 Safety pays+0.49+0.42 to +0.56 · cautious76.4%211 / 276unstable: fewer than 200 needless risks behind a mean cost +0.55β 1.46 · needless risks lost 0.190 on average (187) · missed risks 0.192 (350)
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 BlackwelldecisionLeague 156.654.6 to 58.733.1%213 / 64480.2%733 / 91445.7%213 / 46655.7%359 / 64419.8%181 / 91411.2%72 / 6440.700.60 to 0.81 · risks taken 285 / 644 Risk pays · needless 181 / 914 Safety pays+0.50+0.42 to +0.57 · cautious74.7%213 / 285unstable: fewer than 200 needless risks behind a mean cost +0.30β 1.23 · needless risks lost 0.155 on average (181) · missed risks 0.179 (359)
SubQ Snap PreviewdecisionLeague 155.953.1 to 58.838.4%263 / 68573.4%730 / 99440.6%263 / 64744.1%302 / 68526.6%264 / 99417.5%120 / 6850.770.63 to 0.93 · risks taken 383 / 685 Risk pays · needless 264 / 994 Safety pays+0.24+0.16 to +0.31 · cautious68.7%263 / 383+0.40β 1.37 · needless risks lost 0.189 on average (264) · missed risks 0.201 (302)
GPT-6 LunaDecisions APIdecisionLeague 155.253.2 to 57.121.7%144 / 66388.7%697 / 78650.7%144 / 28470.6%468 / 66311.3%89 / 7867.7%51 / 6630.670.50 to 0.84 · risks taken 195 / 663 Risk pays · needless 89 / 786 Safety pays+0.87+0.79 to +0.96 · cautious73.8%144 / 195unstable: fewer than 200 needless risks behind a mean cost +0.05β 1.03 · needless risks lost 0.179 on average (89) · missed risks 0.205 (468)
Winnow-12BNVIDIA RTX PRO 6000 BlackwelldecisionLeague 154.552.7 to 56.233.6%233 / 69375.3%629 / 83543.1%233 / 54051.8%359 / 69324.7%206 / 83514.6%101 / 6930.640.52 to 0.76 · risks taken 334 / 693 Risk pays · needless 206 / 835 Safety pays+0.36+0.28 to +0.45 · cautious69.8%233 / 334+0.24β 1.17 · needless risks lost 0.211 on average (206) · missed risks 0.218 (359)
Mercury DecidedecisionLeague 153.151.3 to 55.118.4%115 / 62587.8%638 / 72745.6%115 / 25273.9%462 / 62512.2%89 / 7277.7%48 / 6250.520.37 to 0.67 · risks taken 163 / 625 Risk pays · needless 89 / 727 Safety pays+0.90+0.82 to +0.99 · cautious70.6%115 / 163unstable: fewer than 200 needless risks behind a mean cost +0.32β 1.18 · needless risks lost 0.233 on average (89) · missed risks 0.230 (462)
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX SparkchatLeague 152.450.2 to 54.632.2%224 / 69572.6%708 / 97534.5%224 / 65044.9%312 / 69527.4%267 / 97522.9%159 / 6950.730.61 to 0.85 · risks taken 383 / 695 Risk pays · needless 267 / 975 Safety pays+0.24+0.18 to +0.30 · cautious58.5%224 / 383+1.01β 2.09 · needless risks lost 0.243 on average (267) · missed risks 0.163 (312)
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 BlackwelldecisionLeague 148.045.5 to 50.737.7%236 / 62658.4%492 / 84330.7%236 / 76833.4%209 / 62641.6%351 / 84328.9%181 / 6260.640.50 to 0.78 · risks taken 417 / 626 Risk pays · needless 351 / 843 Safety pays−0.11−0.18 to −0.04 · bold56.6%236 / 417+0.72β 1.58 · needless risks lost 0.271 on average (351) · missed risks 0.231 (209)
Perplexity Decider Flash v1.2NVIDIA RTX PRO 6000 BlackwelldecisionLeague 147.745.0 to 50.130.6%192 / 62764.8%558 / 86128.9%192 / 66542.3%265 / 62735.2%303 / 86127.1%170 / 6270.570.46 to 0.70 · risks taken 362 / 627 Risk pays · needless 303 / 861 Safety pays+0.09+0.02 to +0.17 · cautious53.0%192 / 362+0.82β 1.60 · needless risks lost 0.232 on average (303) · missed risks 0.198 (265)
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 BlackwellchatLeague 144.942.6 to 47.628.4%181 / 63761.5%509 / 82827.9%181 / 64848.4%308 / 63738.5%319 / 82823.2%148 / 6370.330.20 to 0.46 · risks taken 329 / 637 Risk pays · needless 319 / 828 Safety pays+0.13+0.05 to +0.21 · cautious55.0%181 / 329+1.71β 1.77 · needless risks lost 0.249 on average (319) · missed risks 0.183 (308)
GNUbg weakprogram · reference row, never ranked · x86-64 build: its noise is shifted, so it overplays gammonsLeague 164.361.1 to 67.566.1%345 / 52262.6%552 / 88242.9%345 / 8059.0%47 / 52237.4%330 / 88224.9%130 / 5221.661.47 to 1.88 · risks taken 475 / 522 Risk pays · needless 330 / 882 Safety pays−0.51−0.62 to −0.41 · bold72.6%345 / 475unstable: fewer than 200 missed risks behind a mean cost +0.47β 2.19 · needless risks lost 0.105 on average (330) · missed risks 0.081 (47)
Randomprogram · reference row, never ranked · a random legal playLeague 123.721.0 to 26.322.7%94 / 41424.7%129 / 52212.9%94 / 72819.1%79 / 41475.3%393 / 52258.2%241 / 4140.190.05 to 0.35 · risks taken 335 / 414 Risk pays · needless 393 / 522 Safety pays−0.78−0.90 to −0.66 · bold28.1%94 / 335unstable: fewer than 200 missed risks behind a mean cost +2.95β 1.75 · needless risks lost 0.377 on average (393) · missed risks 0.271 (79)
Safestprogram · reference row, never ranked · always plays safeLeague 150.050.0 to 50.00.0%0 / 707100.0%827 / 827N/A0 / 0100.0%707 / 7070.0%0 / 8270.0%0 / 7070.040.01 to 0.08 · risks taken 0 / 707 Risk pays · needless 0 / 827 Safety pays+3.21+3.20 to +3.23 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.200 (707)
GNUbg weakprogram · reference row, never ranked · x86-64 build: its noise is shifted, so it overplays gammonsLeague 261.758.7 to 64.664.8%420 / 64858.5%599 / 1,02441.9%420 / 1,00310.8%70 / 64841.5%425 / 1,02424.4%158 / 6481.451.31 to 1.61 · risks taken 578 / 648 Risk pays · needless 425 / 1,024 Safety pays−0.51−0.60 to −0.42 · bold72.7%420 / 578unstable: fewer than 200 missed risks behind a mean cost +0.36β 1.68 · needless risks lost 0.130 on average (425) · missed risks 0.122 (70)
Random (computed)program · reference row, never ranked · computed: every legal play equally likelyLeague 225.925.6 to 26.123.7%2,710.4 / 11,45628.1%5,001.1 / 17,82012.1%2,710.4 / 22,427.916.1%1,847.0 / 11,45671.9%12,818.9 / 17,82060.2%6,898.7 / 11,4560.410.39 to 0.43 · risks taken 9,609.0 / 11,456 Risk pays · needless 12,818.9 / 17,820 Safety pays−0.79−0.79 to −0.78 · bold28.2%2,710.4 / 9,609.0+2.43β 2.69 · needless risks lost 0.381 on average (12,818.9) · missed risks 0.220 (1,847.0)
Safest (computed)program · reference row, never ranked · computed: always the best safe playLeague 250.050.0 to 50.00.0%0 / 11,456100.0%17,820 / 17,820N/A0 / 0100.0%11,456 / 11,4560.0%0 / 17,8200.0%0 / 11,4560.110.10 to 0.11 · risks taken 0 / 11,456 Risk pays · needless 0 / 17,820 Safety pays+3.98+3.97 to +3.98 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.172 (11,456)
GNUbg weakprogram · reference row, never ranked · x86-64 build: its noise is shifted, so it overplays gammonsLeague 360.056.4 to 63.458.1%247 / 42561.8%432 / 69939.4%247 / 62715.3%65 / 42538.2%267 / 69926.6%113 / 4251.321.14 to 1.52 · risks taken 360 / 425 Risk pays · needless 267 / 699 Safety pays−0.36−0.45 to −0.27 · bold68.6%247 / 360unstable: fewer than 200 missed risks behind a mean cost +0.66β 2.38 · needless risks lost 0.132 on average (267) · missed risks 0.091 (65)
Random (computed)program · reference row, never ranked · computed: every legal play equally likelyLeague 326.425.9 to 26.923.4%1,054.8 / 4,50629.4%2,178.6 / 7,41811.7%1,054.8 / 8,984.916.9%760.5 / 4,50670.6%5,239.4 / 7,41859.7%2,690.7 / 4,5060.420.39 to 0.44 · risks taken 3,745.5 / 4,506 Risk pays · needless 5,239.4 / 7,418 Safety pays−0.75−0.77 to −0.73 · bold28.2%1,054.8 / 3,745.5+2.51β 2.84 · needless risks lost 0.398 on average (5,239.4) · missed risks 0.231 (760.5)
Safest (computed)program · reference row, never ranked · computed: always the best safe playLeague 350.050.0 to 50.00.0%0 / 4,506100.0%7,418 / 7,418N/A0 / 0100.0%4,506 / 4,5060.0%0 / 7,4180.0%0 / 4,5060.120.11 to 0.14 · risks taken 0 / 4,506 Risk pays · needless 0 / 7,418 Safety pays+3.76+3.75 to +3.76 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.169 (4,506)

Ability and caution by reasoning setting (UNPAIRED) are in each league's view: League 1, 2 or 3.

Swipe for more columns →

Risk calls at the 0.001 margin, all phases; shares with count / denominator, risk judgement with its 95% interval (games resampled)
Model and settingRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksError rate by risk-call category (PR)Decisions with a trade-offBig risks foundBig needless risksExtra hitting rolls: right / wrong and needlessConfidence on the call: ranking score for right riskAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
Jev 1.13decision model · API59.557.4 to 61.627.0%190 / 70592.1%710 / 77161.7%190 / 30865.0%458 / 7057.9%61 / 7718.1%57 / 70534.4needless risks 2.10 · missed risks 21.41 · wrong risks 3.52 · right risks 2.12 · right safety 2.66 · No trade-off 2.601,476Risk pays 705 · Safety pays 771 · No trade-off 43533.0%143 / 4333.8%29 / 771Too few decisions / Too few decisions190 / 118 decisions0.53308 risks with a probability‡; mean confidence on missed risks 52.2% (458)1.020.82 to 1.26 · risks taken 247 / 705 Risk pays · needless 61 / 771 Safety pays+0.90+0.81 to +1.00 · cautious76.9%190 / 247unstable: fewer than 200 needless risks behind a mean cost −0.21β 0.80 · needless risks lost 0.131 on average (61) · missed risks 0.179 (458)
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release59.357.1 to 61.529.1%192 / 65989.5%603 / 67461.9%192 / 31063.7%420 / 65910.5%71 / 6747.1%47 / 65935.6needless risks 2.94 · missed risks 20.07 · wrong risks 3.16 · right risks 3.52 · right safety 3.53 · No trade-off 2.341,333Risk pays 659 · Safety pays 674 · No trade-off 40338.3%150 / 3925.6%38 / 674Too few decisions / Too few decisions192 / 118 decisions0.54269 risks with a probability‡; 41 split into groups left out; mean confidence on missed risks 72.5% (378)0.900.74 to 1.06 · risks taken 239 / 659 Risk pays · needless 71 / 674 Safety pays+0.80+0.72 to +0.89 · cautious80.3%192 / 239unstable: fewer than 200 needless risks behind a mean cost −0.13β 0.89 · needless risks lost 0.144 on average (71) · missed risks 0.166 (420)
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs58.856.7 to 60.833.0%219 / 66484.6%698 / 82554.8%219 / 40058.9%391 / 66415.4%127 / 8258.1%54 / 66435.0needless risks 5.15 · missed risks 16.67 · wrong risks 2.72 · right risks 1.56 · right safety 5.81 · No trade-off 3.091,489Risk pays 664 · Safety pays 825 · No trade-off 51438.4%139 / 3629.1%75 / 82512.4 / Too few decisions219 / 181 decisions0.60400 risks with a probability‡; mean confidence on missed risks 35.5% (391)0.790.67 to 0.93 · risks taken 273 / 664 Risk pays · needless 127 / 825 Safety pays+0.62+0.54 to +0.71 · cautious80.2%219 / 273unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.18 · needless risks lost 0.162 on average (127) · missed risks 0.171 (391)
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs57.956.0 to 59.933.4%244 / 73182.5%682 / 82752.4%244 / 46656.1%410 / 73117.5%145 / 82710.5%77 / 73136.9needless risks 6.42 · missed risks 16.25 · wrong risks 3.85 · right risks 2.17 · right safety 4.74 · No trade-off 3.511,558Risk pays 731 · Safety pays 827 · No trade-off 47038.9%162 / 41612.2%101 / 82710.1 / 8.9244 / 222 decisions0.55416 risks with a probability‡; 50 split into groups left out; mean confidence on missed risks 42.4% (337)0.780.67 to 0.89 · risks taken 321 / 731 Risk pays · needless 145 / 827 Safety pays+0.54+0.48 to +0.61 · cautious76.0%244 / 321unstable: fewer than 200 needless risks behind a mean cost +0.30β 1.26 · needless risks lost 0.180 on average (145) · missed risks 0.161 (410)
celeris-1-decisiondecision model · API57.855.9 to 59.723.9%163 / 68391.7%749 / 81760.8%163 / 26870.7%483 / 6838.3%68 / 8175.4%37 / 68335.6needless risks 2.30 · missed risks 20.00 · wrong risks 2.07 · right risks 1.37 · right safety 5.50 · No trade-off 4.401,500Risk pays 683 · Safety pays 817 · No trade-off 42232.1%120 / 3744.7%38 / 817Too few decisions / Too few decisions163 / 105 decisions0.52268 risks with a probability‡; mean confidence on missed risks 67.9% (483)0.840.67 to 1.03 · risks taken 200 / 683 Risk pays · needless 68 / 817 Safety pays+0.96+0.88 to +1.05 · cautious81.5%163 / 200unstable: fewer than 200 needless risks behind a mean cost −0.03β 0.97 · needless risks lost 0.130 on average (68) · missed risks 0.159 (483)
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark57.656.1 to 59.023.5%178 / 75691.7%671 / 73263.3%178 / 28170.9%536 / 7568.3%61 / 7325.6%42 / 75645.0needless risks 3.20 · missed risks 26.37 · wrong risks 2.82 · right risks 2.92 · right safety 5.75 · No trade-off 3.981,488Risk pays 756 · Safety pays 732 · No trade-off 51726.3%124 / 4716.3%46 / 732Too few decisions / Too few decisions178 / 103 decisions0.50244 risks with a probability‡; 37 split into groups left out; mean confidence on missed risks 70.5% (477)0.830.70 to 0.98 · risks taken 220 / 756 Risk pays · needless 61 / 732 Safety pays+0.96+0.89 to +1.04 · cautious80.9%178 / 220unstable: fewer than 200 needless risks behind a mean cost +0.04β 1.03 · needless risks lost 0.210 on average (61) · missed risks 0.197 (536)
CygnetNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs57.155.3 to 58.827.3%205 / 75086.8%685 / 78955.9%205 / 36764.9%487 / 75013.2%104 / 7897.7%58 / 75038.6needless risks 4.14 · missed risks 19.98 · wrong risks 3.19 · right risks 2.79 · right safety 4.26 · No trade-off 4.201,539Risk pays 750 · Safety pays 789 · No trade-off 47735.6%159 / 4478.5%67 / 7897.6 / Too few decisions205 / 162 decisions0.56339 risks with a probability‡; 28 split into groups left out; mean confidence on missed risks 65.9% (452)0.730.60 to 0.88 · risks taken 263 / 750 Risk pays · needless 104 / 789 Safety pays+0.75+0.67 to +0.83 · cautious77.9%205 / 263unstable: fewer than 200 needless risks behind a mean cost +0.03β 1.02 · needless risks lost 0.160 on average (104) · missed risks 0.165 (487)
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs56.954.9 to 58.833.1%233 / 70380.7%682 / 84550.3%233 / 46357.3%403 / 70319.3%163 / 8459.5%67 / 70338.8needless risks 8.32 · missed risks 16.99 · wrong risks 3.72 · right risks 2.11 · right safety 3.90 · No trade-off 3.751,548Risk pays 703 · Safety pays 845 · No trade-off 50439.9%160 / 40114.2%120 / 84510.1 / 10.1233 / 230 decisions0.55400 risks with a probability‡; 63 split into groups left out; mean confidence on missed risks 53.3% (340)0.680.56 to 0.81 · risks taken 300 / 703 Risk pays · needless 163 / 845 Safety pays+0.53+0.46 to +0.59 · cautious77.7%233 / 300unstable: fewer than 200 needless risks behind a mean cost +0.55β 1.45 · needless risks lost 0.209 on average (163) · missed risks 0.173 (403)
SubQ Snap Previewdecision model · API56.553.8 to 59.137.9%297 / 78375.1%673 / 89645.9%297 / 64745.8%359 / 78324.9%223 / 89616.2%127 / 78343.9needless risks 10.40 · missed risks 14.35 · wrong risks 7.22 · right risks 3.21 · right safety 5.11 · No trade-off 3.581,679Risk pays 783 · Safety pays 896 · No trade-off 52145.9%201 / 43817.4%156 / 8969.6 / 10.9297 / 350 decisions0.61647 risks with a probability‡; mean confidence on missed risks 44.3% (358)0.780.65 to 0.93 · risks taken 424 / 783 Risk pays · needless 223 / 896 Safety pays+0.29+0.22 to +0.36 · cautious70.0%297 / 424+0.37β 1.33 · needless risks lost 0.205 on average (223) · missed risks 0.176 (359)
GPT-6 LunaDecisions APIdecision model · API55.753.8 to 57.621.6%160 / 74089.7%636 / 70956.3%160 / 28471.5%529 / 74010.3%73 / 7096.9%51 / 74043.3needless risks 3.67 · missed risks 25.05 · wrong risks 3.45 · right risks 2.58 · right safety 4.50 · No trade-off 4.071,449Risk pays 740 · Safety pays 709 · No trade-off 51328.6%123 / 4307.5%53 / 709Too few decisions / Too few decisions160 / 124 decisions0.60284 risks with a probability‡; mean confidence on missed risks 54.6% (529)0.690.54 to 0.87 · risks taken 211 / 740 Risk pays · needless 73 / 709 Safety pays+0.91+0.83 to +1.00 · cautious75.8%160 / 211unstable: fewer than 200 needless risks behind a mean cost +0.02β 1.02 · needless risks lost 0.197 on average (73) · missed risks 0.186 (529)
Winnow-12BNVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs55.252.9 to 57.233.7%263 / 78076.6%573 / 74848.7%263 / 54053.2%415 / 78023.4%175 / 74813.1%102 / 78049.7needless risks 10.05 · missed risks 19.93 · wrong risks 6.06 · right risks 3.45 · right safety 5.37 · No trade-off 4.861,528Risk pays 780 · Safety pays 748 · No trade-off 48739.9%176 / 44116.4%123 / 7489.8 / 10.2263 / 277 decisions0.62518 risks with a probability‡; 22 split into groups left out; mean confidence on missed risks 59.7% (390)0.640.53 to 0.77 · risks taken 365 / 780 Risk pays · needless 175 / 748 Safety pays+0.40+0.32 to +0.49 · cautious72.1%263 / 365unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.15 · needless risks lost 0.231 on average (175) · missed risks 0.194 (415)
Mercury Decidedecision model · API53.451.5 to 55.218.7%129 / 69088.1%583 / 66251.2%129 / 25274.9%517 / 69011.9%79 / 6626.4%44 / 69047.9needless risks 5.59 · missed risks 29.17 · wrong risks 2.52 · right risks 1.67 · right safety 4.44 · No trade-off 4.551,352Risk pays 690 · Safety pays 662 · No trade-off 50121.8%91 / 4177.7%51 / 662Too few decisions / Too few decisions129 / 123 decisions0.63252 risks with a probability‡; mean confidence on missed risks 83.3% (517)0.500.36 to 0.67 · risks taken 173 / 690 Risk pays · needless 79 / 662 Safety pays+0.92+0.84 to +1.02 · cautious74.6%129 / 173unstable: fewer than 200 needless risks behind a mean cost +0.37β 1.20 · needless risks lost 0.262 on average (79) · missed risks 0.209 (517)
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark52.450.3 to 54.531.5%248 / 78873.4%647 / 88238.2%248 / 65047.3%373 / 78826.6%235 / 88221.2%167 / 78847.7needless risks 13.29 · missed risks 12.21 · wrong risks 10.08 · right risks 2.73 · right safety 4.23 · No trade-off 5.151,670Risk pays 788 · Safety pays 882 · No trade-off 52344.8%182 / 40619.2%169 / 88210.5 / 12.3248 / 402 decisions0.52572 risks with a probability‡; 78 split into groups left out; mean confidence on missed risks 100.0% (309)0.690.56 to 0.82 · risks taken 415 / 788 Risk pays · needless 235 / 882 Safety pays+0.28+0.21 to +0.34 · cautious59.8%248 / 415+0.96β 1.93 · needless risks lost 0.248 on average (235) · missed risks 0.144 (373)
Perplexity Decider Flash v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs49.146.5 to 51.631.3%223 / 71266.9%519 / 77633.5%223 / 66542.7%304 / 71233.1%257 / 77626.0%185 / 71259.6needless risks 16.04 · missed risks 13.43 · wrong risks 13.80 · right risks 3.10 · right safety 6.17 · No trade-off 7.111,488Risk pays 712 · Safety pays 776 · No trade-off 51841.9%167 / 39925.3%196 / 77610.7 / 13.4223 / 442 decisions0.56565 risks with a probability‡; 100 split into groups left out; mean confidence on missed risks 32.5% (270)0.620.51 to 0.74 · risks taken 408 / 712 Risk pays · needless 257 / 776 Safety pays+0.13+0.05 to +0.20 · cautious54.7%223 / 408+0.70β 1.54 · needless risks lost 0.250 on average (257) · missed risks 0.177 (304)
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs48.546.1 to 50.937.3%264 / 70859.8%455 / 76134.4%264 / 76834.7%246 / 70840.2%306 / 76128.0%198 / 70865.6needless risks 21.42 · missed risks 12.49 · wrong risks 17.29 · right risks 3.05 · right safety 4.95 · No trade-off 6.351,469Risk pays 708 · Safety pays 761 · No trade-off 53342.8%170 / 39731.0%236 / 76111.2 / 11.9264 / 504 decisions0.53768 risks with a probability‡; mean confidence on missed risks 14.9% (246)0.640.52 to 0.76 · risks taken 462 / 708 Risk pays · needless 306 / 761 Safety pays−0.07−0.14 to −0.01 · bold57.1%264 / 462+0.62β 1.48 · needless risks lost 0.280 on average (306) · missed risks 0.203 (246)
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell45.743.2 to 48.128.8%206 / 71562.7%470 / 75031.8%206 / 64848.5%347 / 71537.3%280 / 75022.7%162 / 71561.5needless risks 18.70 · missed risks 14.79 · wrong risks 13.15 · right risks 2.81 · right safety 5.11 · No trade-off 6.901,465Risk pays 715 · Safety pays 750 · No trade-off 48735.6%147 / 41327.5%206 / 7509.9 / 12.8206 / 442 decisions0.52573 risks with a probability‡; 75 split into groups left out; mean confidence on missed risks 39.2% (316)0.360.23 to 0.49 · risks taken 368 / 715 Risk pays · needless 280 / 750 Safety pays+0.14+0.06 to +0.22 · cautious56.0%206 / 368+1.38β 1.64 · needless risks lost 0.261 on average (280) · missed risks 0.166 (347)
GNUbg weakprogram · CPU · x86-64 build: its noise is shifted, so it overplays gammons67.163.8 to 70.366.6%410 / 61667.6%533 / 78850.9%410 / 80510.7%66 / 61632.4%255 / 78822.7%140 / 61621.4needless risks 8.80 · missed risks 1.22 · wrong risks 5.99 · right risks 2.66 · right safety 1.66 · No trade-off 1.071,404Risk pays 616 · Safety pays 788 · No trade-off 34082.7%254 / 30718.0%142 / 78811.2 / 13.1410 / 395 decisionsN/A1.701.53 to 1.89 · risks taken 550 / 616 Risk pays · needless 255 / 788 Safety pays−0.39−0.48 to −0.30 · bold74.5%410 / 550unstable: fewer than 200 missed risks behind a mean cost +0.51β 2.39 · needless risks lost 0.120 on average (255) · missed risks 0.065 (66)
Randomprogram · CPU24.922.3 to 27.424.8%112 / 45225.0%121 / 48415.4%112 / 72819.2%87 / 45275.0%363 / 48456.0%253 / 452121.1needless risks 54.22 · missed risks 8.20 · wrong risks 38.00 · right risks 3.36 · right safety 3.88 · No trade-off 13.44936Risk pays 452 · Safety pays 484 · No trade-off 40129.2%78 / 26768.0%329 / 484Too few decisions / 16.4112 / 616 decisionsN/A0.190.05 to 0.34 · risks taken 365 / 452 Risk pays · needless 363 / 484 Safety pays−0.77−0.89 to −0.65 · bold30.7%112 / 365unstable: fewer than 200 missed risks behind a mean cost +2.73β 1.70 · needless risks lost 0.399 on average (363) · missed risks 0.252 (87)
Safestprogram · CPU50.050.0 to 50.00.0%0 / 785100.0%749 / 749N/A0 / 0100.0%785 / 7850.0%0 / 7490.0%0 / 78551.6needless risks 0.00 · missed risks 34.95 · wrong risks 0.00 · right risks 0.00 · right safety 8.45 · No trade-off 8.251,534Risk pays 785 · Safety pays 749 · No trade-off 5240.0%0 / 4600.0%0 / 7490 / 00 / 0 decisionsN/A−0.01−0.05 to 0.02 · risks taken 0 / 785 Risk pays · needless 0 / 749 Safety pays+3.22+3.20 to +3.23 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.183 (785)

Hit risk only: being hit. Gammon risk and priming are not in it.

Setups met different positions; risk judgement is not adjusted for how hard they were. Paired comparisons wait for the shared positions.

Big risks found: stake above 0.08. Big needless risks: the chosen play lost more than 0.08.

‡ Risks with a probability: decisions whose plays were split into groups have no single probability and are left out; their count is shown per row. Setups that return no probabilities show N/A. Confidence on the call: ranking score only, no calibration gap (the probability is for the chosen play, not the risk). Definitions

Ability, caution and the SDT reference: A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity. Rates corrected by adding 0.5 to the counts and 1 to the totals; a setup without Risk pays or Safety pays decisions shows N/A. Intervals: 95%, blocks of two seat-swapped games resampled. Definitions

The same at the 0.02 margin (robustness)

Swipe for more columns →

Every 0.001 in the definitions replaced by 0.02
Model and settingRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
DeepSeek-V4.1-Flashoff · API + NVIDIA RTX PRO 6000 Blackwell · first release59.056.7 to 61.330.1%177 / 58887.9%655 / 74557.1%177 / 31062.6%368 / 58812.1%90 / 7457.3%43 / 5880.850.72 to 0.98 · risks taken 220 / 588 Risk pays · needless 90 / 745 Safety pays+0.74+0.66 to +0.83 · cautious80.5%177 / 220unstable: fewer than 200 needless risks behind a mean cost −0.20β 0.84 · needless risks lost 0.124 on average (90) · missed risks 0.187 (368)
Jev 1.13decision model · API58.456.2 to 60.826.8%167 / 62290.0%769 / 85454.2%167 / 30864.1%399 / 62210.0%85 / 8549.0%56 / 6220.920.72 to 1.13 · risks taken 223 / 622 Risk pays · needless 85 / 854 Safety pays+0.82+0.74 to +0.91 · cautious74.9%167 / 223unstable: fewer than 200 needless risks behind a mean cost −0.36β 0.72 · needless risks lost 0.106 on average (85) · missed risks 0.202 (399)
GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs57.955.4 to 60.433.1%192 / 58082.7%752 / 90948.0%192 / 40058.1%337 / 58017.3%157 / 9098.8%51 / 5800.740.59 to 0.89 · risks taken 243 / 580 Risk pays · needless 157 / 909 Safety pays+0.57+0.49 to +0.66 · cautious79.0%192 / 243unstable: fewer than 200 needless risks behind a mean cost +0.15β 1.11 · needless risks lost 0.138 on average (157) · missed risks 0.194 (337)
CygnetNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs57.355.4 to 59.228.2%192 / 68086.4%742 / 85952.3%192 / 36763.2%430 / 68013.6%117 / 8598.5%58 / 6800.760.63 to 0.89 · risks taken 250 / 680 Risk pays · needless 117 / 859 Safety pays+0.72+0.64 to +0.80 · cautious76.8%192 / 250unstable: fewer than 200 needless risks behind a mean cost +0.05β 1.04 · needless risks lost 0.151 on average (117) · missed risks 0.184 (430)
celeris-1-decisiondecision model · API57.355.4 to 59.124.1%141 / 58690.5%827 / 91452.6%141 / 26869.1%405 / 5869.5%87 / 9146.8%40 / 5860.810.65 to 0.98 · risks taken 181 / 586 Risk pays · needless 87 / 914 Safety pays+0.90+0.82 to +0.99 · cautious77.9%141 / 181unstable: fewer than 200 needless risks behind a mean cost −0.10β 0.92 · needless risks lost 0.107 on average (87) · missed risks 0.181 (405)
Qwen3.8-27Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark56.755.1 to 58.222.8%157 / 68890.6%725 / 80055.9%157 / 28170.1%482 / 6889.4%75 / 8007.1%49 / 6880.790.66 to 0.93 · risks taken 206 / 688 Risk pays · needless 75 / 800 Safety pays+0.92+0.85 to +0.99 · cautious76.2%157 / 206unstable: fewer than 200 needless risks behind a mean cost −0.03β 0.98 · needless risks lost 0.180 on average (75) · missed risks 0.215 (482)
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkdecision model · our GPUs56.754.7 to 58.733.7%211 / 62679.7%735 / 92245.6%211 / 46355.9%350 / 62620.3%187 / 92210.4%65 / 6260.680.55 to 0.82 · risks taken 276 / 626 Risk pays · needless 187 / 922 Safety pays+0.49+0.42 to +0.56 · cautious76.4%211 / 276unstable: fewer than 200 needless risks behind a mean cost +0.55β 1.46 · needless risks lost 0.190 on average (187) · missed risks 0.192 (350)
Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs56.654.6 to 58.733.1%213 / 64480.2%733 / 91445.7%213 / 46655.7%359 / 64419.8%181 / 91411.2%72 / 6440.700.60 to 0.81 · risks taken 285 / 644 Risk pays · needless 181 / 914 Safety pays+0.50+0.42 to +0.57 · cautious74.7%213 / 285unstable: fewer than 200 needless risks behind a mean cost +0.30β 1.23 · needless risks lost 0.155 on average (181) · missed risks 0.179 (359)
SubQ Snap Previewdecision model · API55.953.1 to 58.838.4%263 / 68573.4%730 / 99440.6%263 / 64744.1%302 / 68526.6%264 / 99417.5%120 / 6850.770.63 to 0.93 · risks taken 383 / 685 Risk pays · needless 264 / 994 Safety pays+0.24+0.16 to +0.31 · cautious68.7%263 / 383+0.40β 1.37 · needless risks lost 0.189 on average (264) · missed risks 0.201 (302)
GPT-6 LunaDecisions APIdecision model · API55.253.2 to 57.121.7%144 / 66388.7%697 / 78650.7%144 / 28470.6%468 / 66311.3%89 / 7867.7%51 / 6630.670.50 to 0.84 · risks taken 195 / 663 Risk pays · needless 89 / 786 Safety pays+0.87+0.79 to +0.96 · cautious73.8%144 / 195unstable: fewer than 200 needless risks behind a mean cost +0.05β 1.03 · needless risks lost 0.179 on average (89) · missed risks 0.205 (468)
Winnow-12BNVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs54.552.7 to 56.233.6%233 / 69375.3%629 / 83543.1%233 / 54051.8%359 / 69324.7%206 / 83514.6%101 / 6930.640.52 to 0.76 · risks taken 334 / 693 Risk pays · needless 206 / 835 Safety pays+0.36+0.28 to +0.45 · cautious69.8%233 / 334+0.24β 1.17 · needless risks lost 0.211 on average (206) · missed risks 0.218 (359)
Mercury Decidedecision model · API53.151.3 to 55.118.4%115 / 62587.8%638 / 72745.6%115 / 25273.9%462 / 62512.2%89 / 7277.7%48 / 6250.520.37 to 0.67 · risks taken 163 / 625 Risk pays · needless 89 / 727 Safety pays+0.90+0.82 to +0.99 · cautious70.6%115 / 163unstable: fewer than 200 needless risks behind a mean cost +0.32β 1.18 · needless risks lost 0.233 on average (89) · missed risks 0.230 (462)
LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark52.450.2 to 54.632.2%224 / 69572.6%708 / 97534.5%224 / 65044.9%312 / 69527.4%267 / 97522.9%159 / 6950.730.61 to 0.85 · risks taken 383 / 695 Risk pays · needless 267 / 975 Safety pays+0.24+0.18 to +0.30 · cautious58.5%224 / 383+1.01β 2.09 · needless risks lost 0.243 on average (267) · missed risks 0.163 (312)
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs48.045.5 to 50.737.7%236 / 62658.4%492 / 84330.7%236 / 76833.4%209 / 62641.6%351 / 84328.9%181 / 6260.640.50 to 0.78 · risks taken 417 / 626 Risk pays · needless 351 / 843 Safety pays−0.11−0.18 to −0.04 · bold56.6%236 / 417+0.72β 1.58 · needless risks lost 0.271 on average (351) · missed risks 0.231 (209)
Perplexity Decider Flash v1.2NVIDIA RTX PRO 6000 Blackwelldecision model · our GPUs47.745.0 to 50.130.6%192 / 62764.8%558 / 86128.9%192 / 66542.3%265 / 62735.2%303 / 86127.1%170 / 6270.570.46 to 0.70 · risks taken 362 / 627 Risk pays · needless 303 / 861 Safety pays+0.09+0.02 to +0.17 · cautious53.0%192 / 362+0.82β 1.60 · needless risks lost 0.232 on average (303) · missed risks 0.198 (265)
LFM2.5-24B-A2Boff · NVIDIA RTX PRO 6000 Blackwell44.942.6 to 47.628.4%181 / 63761.5%509 / 82827.9%181 / 64848.4%308 / 63738.5%319 / 82823.2%148 / 6370.330.20 to 0.46 · risks taken 329 / 637 Risk pays · needless 319 / 828 Safety pays+0.13+0.05 to +0.21 · cautious55.0%181 / 329+1.71β 1.77 · needless risks lost 0.249 on average (319) · missed risks 0.183 (308)
GNUbg weakprogram · CPU · x86-64 build: its noise is shifted, so it overplays gammons64.361.1 to 67.566.1%345 / 52262.6%552 / 88242.9%345 / 8059.0%47 / 52237.4%330 / 88224.9%130 / 5221.661.47 to 1.88 · risks taken 475 / 522 Risk pays · needless 330 / 882 Safety pays−0.51−0.62 to −0.41 · bold72.6%345 / 475unstable: fewer than 200 missed risks behind a mean cost +0.47β 2.19 · needless risks lost 0.105 on average (330) · missed risks 0.081 (47)
Randomprogram · CPU23.721.0 to 26.322.7%94 / 41424.7%129 / 52212.9%94 / 72819.1%79 / 41475.3%393 / 52258.2%241 / 4140.190.05 to 0.35 · risks taken 335 / 414 Risk pays · needless 393 / 522 Safety pays−0.78−0.90 to −0.66 · bold28.1%94 / 335unstable: fewer than 200 missed risks behind a mean cost +2.95β 1.75 · needless risks lost 0.377 on average (393) · missed risks 0.271 (79)
Safestprogram · CPU50.050.0 to 50.00.0%0 / 707100.0%827 / 827N/A0 / 0100.0%707 / 7070.0%0 / 8270.0%0 / 7070.040.01 to 0.08 · risks taken 0 / 707 Risk pays · needless 0 / 827 Safety pays+3.21+3.20 to +3.23 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.200 (707)

Ability and caution by reasoning setting (UNPAIRED)

Swipe for more columns →

The same model at its lower and higher setting in this league's games: UNPAIRED, the two settings met different positions
ModelAbility d′, 0.001Caution c, 0.001Ability d′, 0.02Caution c, 0.02Median s per decisionOutput tokens per decision (mean)
LFM2.5-24B-A2Boff → on · UNPAIRED0.36 → 0.69change +0.33+0.14 → +0.28change +0.140.33 → 0.73change +0.40+0.13 → +0.24change +0.110.14 → 324.8+324.7 s1 → 21,668

UNPAIRED: league games are not the same positions, so a change here is not a paired effect of the setting. Output tokens include reasoning tokens where the provider counts them there.

The Phase filter does not apply to this table: all phases.

Swipe for more columns →

Risk calls at the 0.001 margin, all phases; shares with count / denominator, risk judgement with its 95% interval (games resampled)
Model and settingRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksError rate by risk-call category (PR)Decisions with a trade-offBig risks foundBig needless risksExtra hitting rolls: right / wrong and needlessConfidence on the call: ranking score for right riskAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first releasereasoning low · letter probabilities · API69.067.3 to 70.945.0%338 / 75193.0%891 / 95875.1%338 / 45049.0%368 / 7517.0%67 / 9586.0%45 / 75119.0needless risks 2.27 · missed risks 8.56 · wrong risks 1.93 · right risks 2.41 · right safety 2.89 · No trade-off 0.931,709Risk pays 751 · Safety pays 958 · No trade-off 47962.8%255 / 4063.4%33 / 9589.1 / Too few decisions338 / 112 decisions0.53371 risks with a probability‡; 79 split into groups left out; mean confidence on missed risks 100.0% (304)1.501.36 to 1.64 · risks taken 383 / 751 Risk pays · needless 67 / 958 Safety pays+0.72+0.63 to +0.81 · cautious88.3%338 / 383unstable: fewer than 200 needless risks behind a mean cost +0.41β 1.86 · needless risks lost 0.148 on average (67) · missed risks 0.102 (368)
DeepSeek-V4.1-Flashlow · API · 1017reasoning low · letter probabilities · API68.566.4 to 70.642.5%321 / 75594.5%872 / 92377.7%321 / 41352.1%393 / 7555.5%51 / 9235.4%41 / 75517.1needless risks 1.48 · missed risks 9.93 · wrong risks 1.57 · right risks 1.68 · right safety 1.63 · No trade-off 0.861,678Risk pays 755 · Safety pays 923 · No trade-off 54455.4%226 / 4082.8%26 / 9239.2 / Too few decisions321 / 92 decisions0.51340 risks with a probability‡; 73 split into groups left out; mean confidence on missed risks 100.0% (337)1.541.39 to 1.70 · risks taken 362 / 755 Risk pays · needless 51 / 923 Safety pays+0.82+0.74 to +0.91 · cautious88.7%321 / 362unstable: fewer than 200 needless risks behind a mean cost +0.22β 1.40 · needless risks lost 0.129 on average (51) · missed risks 0.112 (393)
Claude Sonnet 5.5highreasoning high · one letter · subscription68.266.3 to 70.041.8%308 / 73794.6%841 / 88981.9%308 / 37655.5%409 / 7375.4%48 / 8892.7%20 / 73716.7needless risks 1.12 · missed risks 10.76 · wrong risks 0.64 · right risks 1.70 · right safety 1.58 · No trade-off 0.921,626Risk pays 737 · Safety pays 889 · No trade-off 53856.2%230 / 4091.9%17 / 8898.6 / Too few decisions308 / 68 decisionsN/A1.461.33 to 1.61 · risks taken 328 / 737 Risk pays · needless 48 / 889 Safety pays+0.87+0.80 to +0.95 · cautious93.9%308 / 328unstable: fewer than 200 needless risks behind a mean cost +0.05β 1.07 · needless risks lost 0.101 on average (48) · missed risks 0.114 (409)
Claude Fable 5.1low67.865.6 to 69.941.0%270 / 65894.6%781 / 82677.4%270 / 34953.8%354 / 6585.4%45 / 8265.2%34 / 65817.9needless risks 1.59 · missed risks 8.82 · wrong risks 1.49 · right risks 1.74 · right safety 2.68 · No trade-off 1.611,484Risk pays 658 · Safety pays 826 · No trade-off 52458.2%198 / 3403.4%28 / 8269.3 / Too few decisions270 / 79 decisionsN/A1.501.33 to 1.70 · risks taken 304 / 658 Risk pays · needless 45 / 826 Safety pays+0.85+0.76 to +0.94 · cautious88.8%270 / 304unstable: fewer than 200 needless risks behind a mean cost +0.38β 1.78 · needless risks lost 0.142 on average (45) · missed risks 0.100 (354)
Grok 4.7low67.766.0 to 69.542.3%295 / 69793.0%852 / 91671.4%295 / 41349.9%348 / 6977.0%64 / 9167.7%54 / 69719.1needless risks 1.99 · missed risks 10.40 · wrong risks 2.28 · right risks 1.90 · right safety 2.04 · No trade-off 0.511,613Risk pays 697 · Safety pays 916 · No trade-off 53951.4%201 / 3913.4%31 / 9169.3 / Too few decisions295 / 118 decisionsN/A1.481.34 to 1.63 · risks taken 349 / 697 Risk pays · needless 64 / 916 Safety pays+0.74+0.67 to +0.81 · cautious84.5%295 / 349unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.37 · needless risks lost 0.134 on average (64) · missed risks 0.129 (348)
Gemini 3.8 Flashhighreasoning high · one letter · subscription66.965.2 to 68.536.1%264 / 73197.6%858 / 87987.4%264 / 30261.6%450 / 7312.4%21 / 8792.3%17 / 73114.9needless risks 0.42 · missed risks 10.77 · wrong risks 0.59 · right risks 1.16 · right safety 1.50 · No trade-off 0.501,610Risk pays 731 · Safety pays 879 · No trade-off 55852.9%193 / 3650.5%4 / 8799.2 / Too few decisions264 / 38 decisionsN/A1.681.49 to 1.88 · risks taken 281 / 731 Risk pays · needless 21 / 879 Safety pays+1.13+1.05 to +1.23 · cautious94.0%264 / 281unstable: fewer than 200 needless risks behind a mean cost −0.00β 0.99 · needless risks lost 0.086 on average (21) · missed risks 0.104 (450)
Gemini 3.8 Flashlowreasoning low · one letter · subscription65.163.5 to 66.833.8%214 / 63396.5%877 / 90980.8%214 / 26563.2%400 / 6333.5%32 / 9093.0%19 / 63316.3needless risks 0.80 · missed risks 9.90 · wrong risks 0.74 · right risks 0.79 · right safety 2.82 · No trade-off 1.291,542Risk pays 633 · Safety pays 909 · No trade-off 51149.5%151 / 3051.7%15 / 9099.7 / Too few decisions214 / 51 decisionsN/A1.471.33 to 1.64 · risks taken 233 / 633 Risk pays · needless 32 / 909 Safety pays+1.07+0.99 to +1.17 · cautious91.8%214 / 233unstable: fewer than 200 needless risks behind a mean cost +0.25β 1.45 · needless risks lost 0.102 on average (32) · missed risks 0.102 (400)
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkreasoning on · letter probabilities · our GPUs64.562.9 to 66.231.9%223 / 70097.1%878 / 90482.0%223 / 27264.9%454 / 7002.9%26 / 9043.3%23 / 70019.6needless risks 0.54 · missed risks 12.79 · wrong risks 1.21 · right risks 0.96 · right safety 3.00 · No trade-off 1.081,604Risk pays 700 · Safety pays 904 · No trade-off 51843.7%164 / 3751.2%11 / 90410.3 / Too few decisions223 / 49 decisions0.52225 risks with a probability‡; 47 split into groups left out; mean confidence on missed risks 99.7% (385)1.511.36 to 1.69 · risks taken 246 / 700 Risk pays · needless 26 / 904 Safety pays+1.14+1.05 to +1.23 · cautious90.7%223 / 246unstable: fewer than 200 needless risks behind a mean cost −0.03β 0.95 · needless risks lost 0.088 on average (26) · missed risks 0.120 (454)
d1-ultra64.061.7 to 66.342.0%289 / 68886.0%702 / 81662.7%289 / 46149.6%341 / 68814.0%114 / 8168.4%58 / 68830.3needless risks 3.66 · missed risks 13.25 · wrong risks 3.02 · right risks 2.37 · right safety 4.52 · No trade-off 3.511,504Risk pays 688 · Safety pays 816 · No trade-off 56748.5%180 / 3717.4%60 / 8169.9 / Too few decisions289 / 172 decisions0.64461 risks with a probability‡; mean confidence on missed risks 34.2% (340)1.090.92 to 1.26 · risks taken 347 / 688 Risk pays · needless 114 / 816 Safety pays+0.53+0.46 to +0.61 · cautious83.3%289 / 347unstable: fewer than 200 needless risks behind a mean cost −0.02β 0.98 · needless risks lost 0.133 on average (114) · missed risks 0.161 (341)
Claude Sonnet 5.5lowreasoning low · one letter · subscription63.561.6 to 65.430.9%204 / 66196.2%883 / 91877.6%204 / 26365.5%433 / 6613.8%35 / 9183.6%24 / 66120.1needless risks 1.22 · missed risks 12.96 · wrong risks 0.96 · right risks 1.07 · right safety 2.67 · No trade-off 1.261,579Risk pays 661 · Safety pays 918 · No trade-off 57141.9%145 / 3462.7%25 / 9188.7 / Too few decisions204 / 59 decisionsN/A1.371.19 to 1.58 · risks taken 228 / 661 Risk pays · needless 35 / 918 Safety pays+1.08+0.99 to +1.19 · cautious89.5%204 / 228unstable: fewer than 200 needless risks behind a mean cost +0.35β 1.62 · needless risks lost 0.150 on average (35) · missed risks 0.129 (433)
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkreasoning off · letter probabilities · our GPUs62.660.4 to 65.033.2%194 / 58492.1%732 / 79563.4%194 / 30658.4%341 / 5847.9%63 / 7958.4%49 / 58422.5needless risks 2.06 · missed risks 11.12 · wrong risks 2.12 · right risks 1.55 · right safety 3.07 · No trade-off 2.571,379Risk pays 584 · Safety pays 795 · No trade-off 46844.7%136 / 3044.9%39 / 795Too few decisions / Too few decisions194 / 112 decisions0.60261 risks with a probability‡; 45 split into groups left out; mean confidence on missed risks 97.5% (299)1.201.04 to 1.36 · risks taken 243 / 584 Risk pays · needless 63 / 795 Safety pays+0.81+0.72 to +0.90 · cautious79.8%194 / 243unstable: fewer than 200 needless risks behind a mean cost +0.26β 1.36 · needless risks lost 0.121 on average (63) · missed risks 0.120 (341)
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkreasoning low · letter probabilities · our GPUs62.060.1 to 63.933.8%242 / 71590.2%735 / 81564.0%242 / 37858.3%417 / 7159.8%80 / 8157.8%56 / 71529.0needless risks 2.89 · missed risks 14.82 · wrong risks 2.55 · right risks 1.94 · right safety 3.41 · No trade-off 3.431,530Risk pays 715 · Safety pays 815 · No trade-off 57842.2%174 / 4126.0%49 / 81510.0 / Too few decisions242 / 136 decisions0.56332 risks with a probability‡; 46 split into groups left out; mean confidence on missed risks 100.0% (355)1.080.92 to 1.24 · risks taken 298 / 715 Risk pays · needless 80 / 815 Safety pays+0.75+0.66 to +0.84 · cautious81.2%242 / 298unstable: fewer than 200 needless risks behind a mean cost +0.14β 1.16 · needless risks lost 0.153 on average (80) · missed risks 0.150 (417)
GLM-5.3lowreasoning low · letter probabilities · API61.759.7 to 63.830.7%205 / 66892.7%726 / 78367.7%205 / 30363.2%422 / 6687.3%57 / 7836.1%41 / 66830.9needless risks 1.87 · missed risks 17.40 · wrong risks 2.14 · right risks 1.78 · right safety 3.99 · No trade-off 3.701,451Risk pays 668 · Safety pays 783 · No trade-off 53836.8%141 / 3834.5%35 / 7838.8 / Too few decisions205 / 98 decisions0.52259 risks with a probability‡; 43 split into groups left out; mean confidence on missed risks 91.3% (373)1.120.95 to 1.29 · risks taken 246 / 668 Risk pays · needless 57 / 783 Safety pays+0.89+0.81 to +0.98 · cautious83.3%205 / 246unstable: fewer than 200 needless risks behind a mean cost −0.06β 0.93 · needless risks lost 0.130 on average (57) · missed risks 0.164 (422)
GLiDEAPI60.458.3 to 62.636.9%249 / 67584.0%756 / 90052.4%249 / 47551.0%344 / 67516.0%144 / 90012.1%82 / 67533.2needless risks 7.18 · missed risks 10.97 · wrong risks 4.79 · right risks 2.03 · right safety 4.97 · No trade-off 3.241,575Risk pays 675 · Safety pays 900 · No trade-off 58445.7%169 / 37011.9%107 / 90012.2 / 9.9249 / 226 decisions0.58475 risks with a probability‡; mean confidence on missed risks 39.7% (343)0.970.85 to 1.09 · risks taken 331 / 675 Risk pays · needless 144 / 900 Safety pays+0.51+0.43 to +0.58 · cautious75.2%249 / 331unstable: fewer than 200 needless risks behind a mean cost +0.76β 2.08 · needless risks lost 0.215 on average (144) · missed risks 0.138 (344)
Qwen3.8 Flashoffreasoning off · letter probabilities · API59.557.2 to 61.734.4%243 / 70684.6%613 / 72554.5%243 / 44652.7%372 / 70615.4%112 / 72512.9%91 / 70633.6needless risks 4.64 · missed risks 14.39 · wrong risks 3.96 · right risks 3.46 · right safety 2.96 · No trade-off 4.171,431Risk pays 706 · Safety pays 725 · No trade-off 59642.5%166 / 3919.5%69 / 7257.7 / 9.5243 / 203 decisions0.57386 risks with a probability‡; 60 split into groups left out; mean confidence on missed risks 54.5% (328)0.950.79 to 1.12 · risks taken 334 / 706 Risk pays · needless 112 / 725 Safety pays+0.54+0.46 to +0.62 · cautious72.8%243 / 334unstable: fewer than 200 needless risks behind a mean cost +0.10β 1.10 · needless risks lost 0.168 on average (112) · missed risks 0.157 (372)
Jev 1.1459.357.1 to 61.429.1%184 / 63289.4%657 / 73559.5%184 / 30963.4%401 / 63210.6%78 / 7357.4%47 / 63231.7needless risks 2.84 · missed risks 16.00 · wrong risks 2.91 · right risks 1.72 · right safety 4.37 · No trade-off 3.891,367Risk pays 632 · Safety pays 735 · No trade-off 54533.2%118 / 3556.1%45 / 735Too few decisions / Too few decisions184 / 125 decisions0.49309 risks with a probability‡; mean confidence on missed risks 48.2% (401)0.900.73 to 1.07 · risks taken 231 / 632 Risk pays · needless 78 / 735 Safety pays+0.79+0.73 to +0.86 · cautious79.7%184 / 231unstable: fewer than 200 needless risks behind a mean cost +0.07β 1.06 · needless risks lost 0.139 on average (78) · missed risks 0.153 (401)
DeepSeek-V4.1-Flashoff · API · 1017reasoning off · letter probabilities · API58.556.6 to 60.525.4%181 / 71391.7%683 / 74562.8%181 / 28868.3%487 / 7138.3%62 / 7456.3%45 / 71335.2needless risks 2.49 · missed risks 21.43 · wrong risks 2.50 · right risks 1.61 · right safety 3.75 · No trade-off 3.471,458Risk pays 713 · Safety pays 745 · No trade-off 57633.3%138 / 4155.1%38 / 745Too few decisions / Too few decisions181 / 107 decisions0.58262 risks with a probability‡; 26 split into groups left out; mean confidence on missed risks 81.0% (437)0.900.74 to 1.07 · risks taken 226 / 713 Risk pays · needless 62 / 745 Safety pays+0.93+0.84 to +1.02 · cautious80.1%181 / 226unstable: fewer than 200 needless risks behind a mean cost −0.05β 0.96 · needless risks lost 0.164 on average (62) · missed risks 0.179 (487)
ClefNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark57.555.4 to 59.839.5%263 / 66675.6%603 / 79847.0%263 / 56045.2%301 / 66624.4%195 / 79815.3%102 / 66636.5needless risks 8.85 · missed risks 11.85 · wrong risks 5.48 · right risks 2.26 · right safety 4.18 · No trade-off 3.921,464Risk pays 666 · Safety pays 798 · No trade-off 51447.4%182 / 38415.4%123 / 7989.3 / 10.7263 / 297 decisions0.49552 risks with a probability‡; 8 split into groups left out; mean confidence on missed risks 22.6% (297)0.810.67 to 0.94 · risks taken 365 / 666 Risk pays · needless 195 / 798 Safety pays+0.29+0.20 to +0.37 · cautious72.1%263 / 365unstable: fewer than 200 needless risks behind a mean cost +0.40β 1.38 · needless risks lost 0.180 on average (195) · missed risks 0.156 (301)
GNUbg weakprogram · CPU · x86-64 build: its noise is shifted, so it overplays gammons62.359.5 to 64.862.7%483 / 77061.9%558 / 90248.2%483 / 1,00314.4%111 / 77038.1%344 / 90222.9%176 / 77027.0needless risks 12.05 · missed risks 2.19 · wrong risks 6.03 · right risks 2.93 · right safety 1.97 · No trade-off 1.801,672Risk pays 770 · Safety pays 902 · No trade-off 46482.4%327 / 39723.4%211 / 90211.2 / 13.5483 / 520 decisionsN/A1.361.22 to 1.50 · risks taken 659 / 770 Risk pays · needless 344 / 902 Safety pays−0.38−0.46 to −0.30 · bold73.3%483 / 659unstable: fewer than 200 missed risks behind a mean cost +0.54β 2.08 · needless risks lost 0.150 on average (344) · missed risks 0.084 (111)
Random (computed)reference row, never ranked · computed: every legal play equally likely27.026.7 to 27.225.1%3,300.3 / 13,14028.8%4,655.0 / 16,13614.7%3,300.3 / 22,427.916.7%2,193.1 / 13,14071.2%11,481.0 / 16,13658.2%7,646.6 / 13,140117.9needless risks 57.31 · missed risks 5.40 · wrong risks 40.03 · right risks 3.23 · right safety 3.91 · No trade-off 8.0429,276Risk pays 13,140 · Safety pays 16,136 · No trade-off 10,21231.6%2,248.9 / 7,12763.1%10,184.8 / 16,13615.4 / 17.83,300.3 / 19,127.7 decisionsN/A0.410.39 to 0.42 · risks taken 10,946.9 / 13,140 Risk pays · needless 11,481.0 / 16,136 Safety pays−0.76−0.77 to −0.75 · bold30.1%3,300.3 / 10,946.9+2.23β 2.49 · needless risks lost 0.394 on average (11,481.0) · missed risks 0.195 (2,193.1)
Safest (computed)reference row, never ranked · computed: always the best safe play50.050.0 to 50.00.0%0 / 13,140100.0%16,136 / 16,136N/A0 / 0100.0%13,140 / 13,1400.0%0 / 16,1360.0%0 / 13,14025.2needless risks 0.00 · missed risks 25.23 · wrong risks 0.00 · right risks 0.00 · right safety 0.00 · No trade-off 0.0029,276Risk pays 13,140 · Safety pays 16,136 · No trade-off 10,2120.0%0 / 7,1270.0%0 / 16,1360 / 00 / 0 decisionsN/A0.050.04 to 0.06 · risks taken 0 / 13,140 Risk pays · needless 0 / 16,136 Safety pays+3.98+3.98 to +3.98 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.152 (13,140)

Hit risk only: being hit. Gammon risk and priming are not in it.

Setups met different positions; risk judgement is not adjusted for how hard they were. Paired comparisons wait for the shared positions.

Big risks found: stake above 0.08. Big needless risks: the chosen play lost more than 0.08.

‡ Risks with a probability: decisions whose plays were split into groups have no single probability and are left out; their count is shown per row. Setups that return no probabilities show N/A. Confidence on the call: ranking score only, no calibration gap (the probability is for the chosen play, not the risk). Definitions

Ability, caution and the SDT reference: A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity. Rates corrected by adding 0.5 to the counts and 1 to the totals; a setup without Risk pays or Safety pays decisions shows N/A. Intervals: 95%, blocks of two seat-swapped games resampled. Definitions

The same at the 0.02 margin (robustness)

Swipe for more columns →

Every 0.001 in the definitions replaced by 0.02
Model and settingRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first releasereasoning low · letter probabilities · API70.268.2 to 72.249.2%315 / 64091.2%975 / 1,06970.0%315 / 45044.4%284 / 6408.8%94 / 1,0696.4%41 / 6401.491.36 to 1.63 · risks taken 356 / 640 Risk pays · needless 94 / 1,069 Safety pays+0.60+0.53 to +0.68 · cautious88.5%315 / 356unstable: fewer than 200 needless risks behind a mean cost +0.33β 1.64 · needless risks lost 0.125 on average (94) · missed risks 0.127 (284)
Claude Sonnet 5.5highreasoning high · one letter · subscription69.067.0 to 71.144.0%292 / 66494.0%904 / 96277.7%292 / 37652.1%346 / 6646.0%58 / 9623.9%26 / 6641.501.37 to 1.65 · risks taken 318 / 664 Risk pays · needless 58 / 962 Safety pays+0.80+0.72 to +0.89 · cautious91.8%292 / 318unstable: fewer than 200 needless risks behind a mean cost −0.05β 0.93 · needless risks lost 0.084 on average (58) · missed risks 0.132 (346)
DeepSeek-V4.1-Flashlow · API · 1017reasoning low · letter probabilities · API68.966.5 to 71.344.8%293 / 65493.0%952 / 1,02470.9%293 / 41347.9%313 / 6547.0%72 / 1,0247.3%48 / 6541.521.36 to 1.70 · risks taken 341 / 654 Risk pays · needless 72 / 1,024 Safety pays+0.71+0.63 to +0.79 · cautious85.9%293 / 341unstable: fewer than 200 needless risks behind a mean cost +0.08β 1.13 · needless risks lost 0.097 on average (72) · missed risks 0.135 (313)
Claude Fable 5.1low68.566.3 to 70.843.7%248 / 56793.2%855 / 91771.1%248 / 34949.4%280 / 5676.8%62 / 9176.9%39 / 5671.511.35 to 1.68 · risks taken 287 / 567 Risk pays · needless 62 / 917 Safety pays+0.74+0.65 to +0.83 · cautious86.4%248 / 287unstable: fewer than 200 needless risks behind a mean cost +0.33β 1.65 · needless risks lost 0.122 on average (62) · missed risks 0.120 (280)
Gemini 3.8 Flashhighreasoning high · one letter · subscription67.665.5 to 69.639.2%239 / 61096.1%961 / 1,00079.1%239 / 30256.9%347 / 6103.9%39 / 1,0003.9%24 / 6101.581.40 to 1.79 · risks taken 263 / 610 Risk pays · needless 39 / 1,000 Safety pays+0.97+0.89 to +1.05 · cautious90.9%239 / 263unstable: fewer than 200 needless risks behind a mean cost −0.29β 0.63 · needless risks lost 0.049 on average (39) · missed risks 0.128 (347)
Grok 4.7low67.064.9 to 69.142.8%263 / 61491.1%910 / 99963.7%263 / 41347.2%290 / 6148.9%89 / 9999.9%61 / 6141.411.26 to 1.57 · risks taken 324 / 614 Risk pays · needless 89 / 999 Safety pays+0.64+0.57 to +0.71 · cautious81.2%263 / 324unstable: fewer than 200 needless risks behind a mean cost +0.08β 1.13 · needless risks lost 0.104 on average (89) · missed risks 0.150 (290)
Gemini 3.8 Flashlowreasoning low · one letter · subscription66.864.9 to 68.837.6%203 / 54096.0%962 / 1,00276.6%203 / 26558.3%315 / 5404.0%40 / 1,0024.1%22 / 5401.541.38 to 1.71 · risks taken 225 / 540 Risk pays · needless 40 / 1,002 Safety pays+0.98+0.90 to +1.07 · cautious90.2%203 / 225unstable: fewer than 200 needless risks behind a mean cost +0.22β 1.40 · needless risks lost 0.093 on average (40) · missed risks 0.124 (315)
Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkreasoning on · letter probabilities · our GPUs65.663.8 to 67.534.7%213 / 61396.5%956 / 99178.3%213 / 27261.3%376 / 6133.5%35 / 9913.9%24 / 6131.511.35 to 1.68 · risks taken 237 / 613 Risk pays · needless 35 / 991 Safety pays+1.04+0.96 to +1.14 · cautious89.9%213 / 237unstable: fewer than 200 needless risks behind a mean cost +0.07β 1.11 · needless risks lost 0.095 on average (35) · missed risks 0.139 (376)
Claude Sonnet 5.5lowreasoning low · one letter · subscription64.262.1 to 66.233.1%190 / 57495.2%957 / 1,00572.2%190 / 26362.5%359 / 5744.8%48 / 1,0054.4%25 / 5741.341.18 to 1.53 · risks taken 215 / 574 Risk pays · needless 48 / 1,005 Safety pays+0.99+0.91 to +1.09 · cautious88.4%190 / 215unstable: fewer than 200 needless risks behind a mean cost +0.21β 1.33 · needless risks lost 0.115 on average (48) · missed risks 0.151 (359)
Gemma 4 31Boff · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkreasoning off · letter probabilities · our GPUs63.261.0 to 65.635.5%182 / 51390.9%787 / 86659.5%182 / 30655.8%286 / 5139.1%79 / 8668.8%45 / 5131.191.03 to 1.35 · risks taken 227 / 513 Risk pays · needless 79 / 866 Safety pays+0.74+0.66 to +0.82 · cautious80.2%182 / 227unstable: fewer than 200 needless risks behind a mean cost +0.22β 1.30 · needless risks lost 0.106 on average (79) · missed risks 0.138 (286)
d1-ultra63.060.8 to 65.341.8%254 / 60784.2%755 / 89755.1%254 / 46147.4%288 / 60715.8%142 / 89710.7%65 / 6071.060.91 to 1.22 · risks taken 319 / 607 Risk pays · needless 142 / 897 Safety pays+0.47+0.40 to +0.54 · cautious79.6%254 / 319unstable: fewer than 200 needless risks behind a mean cost −0.02β 0.98 · needless risks lost 0.122 on average (142) · missed risks 0.185 (288)
GLM-5.3lowreasoning low · letter probabilities · API61.659.5 to 63.931.8%184 / 57991.4%797 / 87260.7%184 / 30360.6%351 / 5798.6%75 / 8727.6%44 / 5791.090.91 to 1.27 · risks taken 228 / 579 Risk pays · needless 75 / 872 Safety pays+0.82+0.74 to +0.89 · cautious80.7%184 / 228unstable: fewer than 200 needless risks behind a mean cost −0.12β 0.88 · needless risks lost 0.113 on average (75) · missed risks 0.193 (351)
Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkreasoning low · letter probabilities · our GPUs61.259.2 to 63.234.4%217 / 63087.9%791 / 90057.4%217 / 37857.3%361 / 63012.1%109 / 9008.3%52 / 6300.980.83 to 1.15 · risks taken 269 / 630 Risk pays · needless 109 / 900 Safety pays+0.68+0.59 to +0.77 · cautious80.7%217 / 269unstable: fewer than 200 needless risks behind a mean cost +0.11β 1.11 · needless risks lost 0.131 on average (109) · missed risks 0.169 (361)
GLiDEAPI60.558.2 to 62.638.4%228 / 59382.5%810 / 98248.0%228 / 47548.9%290 / 59317.5%172 / 98212.6%75 / 5930.960.86 to 1.07 · risks taken 303 / 593 Risk pays · needless 172 / 982 Safety pays+0.45+0.38 to +0.53 · cautious75.2%228 / 303unstable: fewer than 200 needless risks behind a mean cost +0.73β 2.01 · needless risks lost 0.190 on average (172) · missed risks 0.157 (290)
Qwen3.8 Flashoffreasoning off · letter probabilities · API59.357.1 to 61.535.5%222 / 62583.0%669 / 80649.8%222 / 44650.6%316 / 62517.0%137 / 80613.9%87 / 6250.940.79 to 1.08 · risks taken 309 / 625 Risk pays · needless 137 / 806 Safety pays+0.48+0.41 to +0.56 · cautious71.8%222 / 309unstable: fewer than 200 needless risks behind a mean cost +0.06β 1.05 · needless risks lost 0.147 on average (137) · missed risks 0.180 (316)
Jev 1.1459.257.0 to 61.429.9%167 / 55988.5%715 / 80854.0%167 / 30961.4%343 / 55911.5%93 / 8088.8%49 / 5590.910.75 to 1.08 · risks taken 216 / 559 Risk pays · needless 93 / 808 Safety pays+0.74+0.68 to +0.81 · cautious77.3%167 / 216unstable: fewer than 200 needless risks behind a mean cost −0.00β 1.00 · needless risks lost 0.121 on average (93) · missed risks 0.174 (343)
DeepSeek-V4.1-Flashoff · API · 1017reasoning off · letter probabilities · API58.856.7 to 60.926.5%168 / 63591.1%750 / 82358.3%168 / 28866.1%420 / 6358.9%73 / 8237.4%47 / 6350.930.77 to 1.10 · risks taken 215 / 635 Risk pays · needless 73 / 823 Safety pays+0.88+0.79 to +0.97 · cautious78.1%168 / 215unstable: fewer than 200 needless risks behind a mean cost −0.09β 0.92 · needless risks lost 0.145 on average (73) · missed risks 0.204 (420)
ClefNVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark57.655.2 to 60.040.6%240 / 59174.6%651 / 87342.9%240 / 56042.8%253 / 59125.4%222 / 87316.6%98 / 5910.840.70 to 0.99 · risks taken 338 / 591 Risk pays · needless 222 / 873 Safety pays+0.24+0.15 to +0.32 · cautious71.0%240 / 338+0.39β 1.39 · needless risks lost 0.168 on average (222) · missed risks 0.178 (253)
GNUbg weakprogram · CPU · x86-64 build: its noise is shifted, so it overplays gammons61.758.7 to 64.664.8%420 / 64858.5%599 / 1,02441.9%420 / 1,00310.8%70 / 64841.5%425 / 1,02424.4%158 / 6481.451.31 to 1.61 · risks taken 578 / 648 Risk pays · needless 425 / 1,024 Safety pays−0.51−0.60 to −0.42 · bold72.7%420 / 578unstable: fewer than 200 missed risks behind a mean cost +0.36β 1.68 · needless risks lost 0.130 on average (425) · missed risks 0.122 (70)
Random (computed)reference row, never ranked · computed: every legal play equally likely25.925.6 to 26.123.7%2,710.4 / 11,45628.1%5,001.1 / 17,82012.1%2,710.4 / 22,427.916.1%1,847.0 / 11,45671.9%12,818.9 / 17,82060.2%6,898.7 / 11,4560.410.39 to 0.43 · risks taken 9,609.0 / 11,456 Risk pays · needless 12,818.9 / 17,820 Safety pays−0.79−0.79 to −0.78 · bold28.2%2,710.4 / 9,609.0+2.43β 2.69 · needless risks lost 0.381 on average (12,818.9) · missed risks 0.220 (1,847.0)
Safest (computed)reference row, never ranked · computed: always the best safe play50.050.0 to 50.00.0%0 / 11,456100.0%17,820 / 17,820N/A0 / 0100.0%11,456 / 11,4560.0%0 / 17,8200.0%0 / 11,4560.110.10 to 0.11 · risks taken 0 / 11,456 Risk pays · needless 0 / 17,820 Safety pays+3.98+3.97 to +3.98 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.172 (11,456)

Ability and caution by reasoning setting (UNPAIRED)

Swipe for more columns →

The same model at its lower and higher setting in this league's games: UNPAIRED, the two settings met different positions
ModelAbility d′, 0.001Caution c, 0.001Ability d′, 0.02Caution c, 0.02Median s per decisionOutput tokens per decision (mean)
Claude Sonnet 5.5low → high · UNPAIRED1.37 → 1.46change +0.10+1.08 → +0.87change −0.211.34 → 1.50change +0.15+0.99 → +0.80change −0.191.51 → 6.63+5.1 s238 → 852
DeepSeek-V4.1-Flash1017off → low · UNPAIRED0.90 → 1.54change +0.64+0.93 → +0.82change −0.110.93 → 1.52change +0.59+0.88 → +0.71change −0.170.84 → 22.3+21.5 s1 → 7,173
Gemini 3.8 Flashlow → high · UNPAIRED1.47 → 1.68change +0.21+1.07 → +1.13change +0.061.54 → 1.58change +0.05+0.98 → +0.97change −0.012.16 → 9.59+7.4 s237 → 5,092
Gemma 4 31Boff → on · UNPAIRED1.20 → 1.51change +0.32+0.81 → +1.14change +0.331.19 → 1.51change +0.33+0.74 → +1.04change +0.310.31 → 120.6+120.3 s1 → 5,599

UNPAIRED: league games are not the same positions, so a change here is not a paired effect of the setting. Output tokens include reasoning tokens where the provider counts them there.

The Phase filter does not apply to this table: all phases.

Swipe for more columns →

Risk calls at the 0.001 margin, all phases; shares with count / denominator, risk judgement with its 95% interval (games resampled)
Model and settingRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksError rate by risk-call category (PR)Decisions with a trade-offBig risks foundBig needless risksExtra hitting rolls: right / wrong and needlessConfidence on the call: ranking score for right riskAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
GPT-6 Astralowreasoning low · one letter · subscription75.873.1 to 78.557.2%226 / 39594.4%486 / 51584.0%226 / 26939.2%155 / 3955.6%29 / 5153.5%14 / 39511.8needless risks 0.64 · missed risks 6.63 · wrong risks 0.69 · right risks 1.81 · right safety 1.03 · No trade-off 1.00910Risk pays 395 · Safety pays 515 · No trade-off 34168.9%153 / 2221.6%8 / 5159.7 / Too few decisions226 / 43 decisionsN/A1.851.63 to 2.14 · risks taken 240 / 395 Risk pays · needless 29 / 515 Safety pays+0.65+0.57 to +0.76 · cautious94.2%226 / 240unstable: fewer than 200 needless and missed risks behind a mean cost −0.21β 0.68 · needless risks lost 0.056 on average (29) · missed risks 0.107 (155)
GPT-6 Astramediumreasoning medium · one letter · subscription74.771.9 to 77.354.9%245 / 44694.4%542 / 57484.8%245 / 28942.4%189 / 4465.6%32 / 5742.7%12 / 44611.1needless risks 0.92 · missed risks 6.28 · wrong risks 0.34 · right risks 1.58 · right safety 1.31 · No trade-off 0.671,020Risk pays 446 · Safety pays 574 · No trade-off 34270.5%170 / 2411.6%9 / 5748.7 / Too few decisions245 / 44 decisionsN/A1.781.59 to 2.00 · risks taken 257 / 446 Risk pays · needless 32 / 574 Safety pays+0.70+0.60 to +0.82 · cautious95.3%245 / 257unstable: fewer than 200 needless and missed risks behind a mean cost +0.06β 1.11 · needless risks lost 0.078 on average (32) · missed risks 0.091 (189)
Gemini 4 Argonhighreasoning high · one letter · subscription74.172.0 to 76.150.8%224 / 44197.4%484 / 49788.5%224 / 25345.6%201 / 4412.6%13 / 4973.6%16 / 44112.6needless risks 0.18 · missed risks 8.05 · wrong risks 0.55 · right risks 1.70 · right safety 1.34 · No trade-off 0.75938Risk pays 441 · Safety pays 497 · No trade-off 39466.4%148 / 2230.4%2 / 49711.0 / Too few decisions224 / 29 decisionsN/A2.041.82 to 2.36 · risks taken 240 / 441 Risk pays · needless 13 / 497 Safety pays+0.91+0.77 to +1.09 · cautious93.3%224 / 240unstable: fewer than 200 needless risks behind a mean cost −0.47β 0.38 · needless risks lost 0.036 on average (13) · missed risks 0.107 (201)
GPT-6.1 Sollow72.769.9 to 75.551.7%198 / 38393.8%602 / 64278.0%198 / 25444.1%169 / 3836.2%40 / 6424.2%16 / 38312.1needless risks 1.00 · missed risks 6.27 · wrong risks 0.83 · right risks 1.62 · right safety 1.78 · No trade-off 0.621,025Risk pays 383 · Safety pays 642 · No trade-off 31965.0%134 / 2061.6%10 / 642Too few decisions / Too few decisions198 / 56 decisionsN/A1.681.51 to 1.88 · risks taken 214 / 383 Risk pays · needless 40 / 642 Safety pays+0.69+0.58 to +0.83 · cautious92.5%198 / 214unstable: fewer than 200 needless and missed risks behind a mean cost +0.07β 1.13 · needless risks lost 0.067 on average (40) · missed risks 0.100 (169)
Claude Fable 5.1highreasoning high · one letter · subscription72.369.8 to 74.750.7%215 / 42494.0%529 / 56377.1%215 / 27942.2%179 / 4246.0%34 / 5637.1%30 / 42414.0needless risks 1.38 · missed risks 6.65 · wrong risks 1.53 · right risks 1.48 · right safety 2.23 · No trade-off 0.68987Risk pays 424 · Safety pays 563 · No trade-off 29965.5%144 / 2202.7%15 / 5639.7 / Too few decisions215 / 64 decisionsN/A1.741.56 to 1.96 · risks taken 245 / 424 Risk pays · needless 34 / 563 Safety pays+0.67+0.55 to +0.81 · cautious87.8%215 / 245unstable: fewer than 200 needless and missed risks behind a mean cost +0.22β 1.46 · needless risks lost 0.105 on average (34) · missed risks 0.096 (179)
GPT-6.1 Solhigh72.269.5 to 74.950.6%212 / 41993.7%568 / 60678.2%212 / 27144.4%186 / 4196.3%38 / 6065.0%21 / 4199.9needless risks 0.89 · missed risks 4.75 · wrong risks 0.90 · right risks 1.27 · right safety 1.53 · No trade-off 0.561,025Risk pays 419 · Safety pays 606 · No trade-off 36471.2%148 / 2081.7%10 / 60610.8 / Too few decisions212 / 59 decisionsN/A1.671.42 to 1.94 · risks taken 233 / 419 Risk pays · needless 38 / 606 Safety pays+0.69+0.60 to +0.79 · cautious91.0%212 / 233unstable: fewer than 200 needless and missed risks behind a mean cost +0.17β 1.32 · needless risks lost 0.065 on average (38) · missed risks 0.071 (186)
Claude Opus 5.5highreasoning high · one letter · subscription70.067.3 to 72.643.4%199 / 45996.6%590 / 61181.6%199 / 24451.4%236 / 4593.4%21 / 6115.2%24 / 45914.4needless risks 0.62 · missed risks 8.96 · wrong risks 1.23 · right risks 1.26 · right safety 1.55 · No trade-off 0.811,070Risk pays 459 · Safety pays 611 · No trade-off 40756.8%146 / 2571.5%9 / 611Too few decisions / Too few decisions199 / 45 decisionsN/A1.771.53 to 2.05 · risks taken 223 / 459 Risk pays · needless 21 / 611 Safety pays+0.92+0.79 to +1.08 · cautious89.2%199 / 223unstable: fewer than 200 needless risks behind a mean cost +0.02β 1.04 · needless risks lost 0.088 on average (21) · missed risks 0.112 (236)
Gemini 4 Argonlowreasoning low · one letter · subscription69.266.2 to 72.347.8%209 / 43790.7%543 / 59971.6%209 / 29246.0%201 / 4379.3%56 / 5996.2%27 / 43716.5needless risks 2.37 · missed risks 8.03 · wrong risks 1.23 · right risks 1.35 · right safety 2.53 · No trade-off 0.941,036Risk pays 437 · Safety pays 599 · No trade-off 44859.1%137 / 2324.0%24 / 59910.8 / Too few decisions209 / 83 decisionsN/A1.421.22 to 1.64 · risks taken 236 / 437 Risk pays · needless 56 / 599 Safety pays+0.61+0.52 to +0.70 · cautious88.6%209 / 236unstable: fewer than 200 needless risks behind a mean cost +0.26β 1.46 · needless risks lost 0.126 on average (56) · missed risks 0.119 (201)
Grok 4.7highreasoning high · one letter · subscription68.566.1 to 71.144.0%190 / 43293.0%454 / 48876.6%190 / 24850.5%218 / 4327.0%34 / 4885.6%24 / 43218.8needless risks 1.31 · missed risks 9.66 · wrong risks 2.39 · right risks 3.33 · right safety 1.53 · No trade-off 0.56920Risk pays 432 · Safety pays 488 · No trade-off 41753.8%129 / 2402.7%13 / 488Too few decisions / Too few decisions190 / 58 decisionsN/A1.461.26 to 1.70 · risks taken 214 / 432 Risk pays · needless 34 / 488 Safety pays+0.74+0.64 to +0.86 · cautious88.8%190 / 214unstable: fewer than 200 needless risks behind a mean cost −0.01β 0.99 · needless risks lost 0.103 on average (34) · missed risks 0.118 (218)
Claude Opus 5.5lowreasoning low · one letter · subscription66.964.7 to 69.240.1%187 / 46693.8%541 / 57777.3%187 / 24255.8%260 / 4666.2%36 / 5774.1%19 / 46615.4needless risks 1.67 · missed risks 8.98 · wrong risks 1.10 · right risks 1.00 · right safety 2.14 · No trade-off 0.551,043Risk pays 466 · Safety pays 577 · No trade-off 36853.0%131 / 2473.5%20 / 577Too few decisions / Too few decisions187 / 55 decisionsN/A1.381.17 to 1.62 · risks taken 206 / 466 Risk pays · needless 36 / 577 Safety pays+0.84+0.72 to +0.97 · cautious90.8%187 / 206unstable: fewer than 200 needless risks behind a mean cost +0.37β 1.66 · needless risks lost 0.131 on average (36) · missed risks 0.097 (260)
Kimi K3.1 Prooffreasoning off · one letter · API65.963.2 to 68.942.4%153 / 36189.5%416 / 46565.7%153 / 23349.0%177 / 36110.5%49 / 4658.6%31 / 36128.0needless risks 4.62 · missed risks 10.08 · wrong risks 3.16 · right risks 2.22 · right safety 4.08 · No trade-off 3.89826Risk pays 361 · Safety pays 465 · No trade-off 34151.4%109 / 2126.7%31 / 465Too few decisions / Too few decisions153 / 80 decisionsN/A1.271.11 to 1.45 · risks taken 184 / 361 Risk pays · needless 49 / 465 Safety pays+0.61+0.51 to +0.72 · cautious83.2%153 / 184unstable: fewer than 200 needless and missed risks behind a mean cost +0.60β 2.13 · needless risks lost 0.220 on average (49) · missed risks 0.133 (177)
GNUbg weakprogram · CPU · x86-64 build: its noise is shifted, so it overplays gammons62.158.2 to 65.958.8%288 / 49065.5%415 / 63445.9%288 / 62716.7%82 / 49034.5%219 / 63424.5%120 / 49028.1needless risks 11.03 · missed risks 2.16 · wrong risks 6.12 · right risks 4.03 · right safety 2.52 · No trade-off 2.221,124Risk pays 490 · Safety pays 634 · No trade-off 35582.8%202 / 24421.5%136 / 63412.5 / 14.5288 / 339 decisionsN/A1.361.16 to 1.57 · risks taken 408 / 490 Risk pays · needless 219 / 634 Safety pays−0.28−0.36 to −0.20 · bold70.6%288 / 408unstable: fewer than 200 missed risks behind a mean cost +0.67β 2.47 · needless risks lost 0.149 on average (219) · missed risks 0.078 (82)
Random (computed)reference row, never ranked · computed: every legal play equally likely27.627.1 to 28.025.1%1,293.3 / 5,15330.1%2,036.5 / 6,77114.4%1,293.3 / 8,984.917.5%902.5 / 5,15369.9%4,734.5 / 6,77157.4%2,957.2 / 5,153117.6needless risks 59.90 · missed risks 5.67 · wrong risks 36.45 · right risks 3.08 · right safety 4.60 · No trade-off 7.9411,924Risk pays 5,153 · Safety pays 6,771 · No trade-off 4,39531.4%862.9 / 2,75262.8%4,253.1 / 6,77115.1 / 17.71,293.3 / 7,691.6 decisionsN/A0.410.39 to 0.44 · risks taken 4,250.5 / 5,153 Risk pays · needless 4,734.5 / 6,771 Safety pays−0.73−0.74 to −0.71 · bold30.4%1,293.3 / 4,250.5+2.36β 2.65 · needless risks lost 0.413 on average (4,734.5) · missed risks 0.205 (902.5)
Safest (computed)reference row, never ranked · computed: always the best safe play50.050.0 to 50.00.0%0 / 5,153100.0%6,771 / 6,771N/A0 / 0100.0%5,153 / 5,1530.0%0 / 6,7710.0%0 / 5,15323.5needless risks 0.00 · missed risks 23.47 · wrong risks 0.00 · right risks 0.00 · right safety 0.00 · No trade-off 0.0011,924Risk pays 5,153 · Safety pays 6,771 · No trade-off 4,3950.0%0 / 2,7520.0%0 / 6,7710 / 00 / 0 decisionsN/A0.070.06 to 0.08 · risks taken 0 / 5,153 Risk pays · needless 0 / 6,771 Safety pays+3.76+3.75 to +3.77 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.149 (5,153)

Hit risk only: being hit. Gammon risk and priming are not in it.

Setups met different positions; risk judgement is not adjusted for how hard they were. Paired comparisons wait for the shared positions.

Big risks found: stake above 0.08. Big needless risks: the chosen play lost more than 0.08.

‡ Risks with a probability: decisions whose plays were split into groups have no single probability and are left out; their count is shown per row. Setups that return no probabilities show N/A. Confidence on the call: ranking score only, no calibration gap (the probability is for the chosen play, not the risk). Definitions

Ability, caution and the SDT reference: A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity. Rates corrected by adding 0.5 to the counts and 1 to the totals; a setup without Risk pays or Safety pays decisions shows N/A. Intervals: 95%, blocks of two seat-swapped games resampled. Definitions

The same at the 0.02 margin (robustness)

Swipe for more columns →

Every 0.001 in the definitions replaced by 0.02
Model and settingRisk judgementRisks foundRight safetyGood risksMissed risksNeedless risksWrong risksAbility d′Caution c (+ cautious, − bold)Aim: right ÷ (right + wrong risks)Cost-derived SDT reference c*
Gemini 4 Argonhighreasoning high · one letter · subscription75.973.4 to 78.156.4%215 / 38195.3%531 / 55785.0%215 / 25340.4%154 / 3814.7%26 / 5573.1%12 / 3811.911.73 to 2.11 · risks taken 227 / 381 Risk pays · needless 26 / 557 Safety pays+0.71+0.59 to +0.84 · cautious94.7%215 / 227unstable: fewer than 200 needless and missed risks behind a mean cost −0.49β 0.40 · needless risks lost 0.036 on average (26) · missed risks 0.132 (154)
GPT-6 Astralowreasoning low · one letter · subscription74.971.9 to 77.958.1%209 / 36091.8%505 / 55077.7%209 / 26937.8%136 / 3608.2%45 / 5504.2%15 / 3601.701.52 to 1.88 · risks taken 224 / 360 Risk pays · needless 45 / 550 Safety pays+0.54+0.47 to +0.62 · cautious93.3%209 / 224unstable: fewer than 200 needless and missed risks behind a mean cost −0.44β 0.47 · needless risks lost 0.037 on average (45) · missed risks 0.118 (136)
GPT-6 Astramediumreasoning medium · one letter · subscription74.972.4 to 77.356.5%225 / 39893.2%580 / 62277.9%225 / 28937.9%151 / 3986.8%42 / 6225.5%22 / 3981.801.62 to 1.99 · risks taken 247 / 398 Risk pays · needless 42 / 622 Safety pays+0.59+0.50 to +0.70 · cautious91.1%225 / 247unstable: fewer than 200 needless and missed risks behind a mean cost −0.09β 0.85 · needless risks lost 0.059 on average (42) · missed risks 0.109 (151)
GPT-6.1 Sollow74.270.9 to 77.456.2%182 / 32492.2%646 / 70171.7%182 / 25438.6%125 / 3247.8%55 / 7015.2%17 / 3241.701.49 to 1.90 · risks taken 199 / 324 Risk pays · needless 55 / 701 Safety pays+0.56+0.46 to +0.68 · cautious91.5%182 / 199unstable: fewer than 200 needless and missed risks behind a mean cost −0.06β 0.90 · needless risks lost 0.054 on average (55) · missed risks 0.129 (125)
GPT-6.1 Solhigh72.469.7 to 75.152.4%188 / 35992.5%616 / 66669.4%188 / 27138.4%138 / 3597.5%50 / 6669.2%33 / 3591.731.52 to 1.96 · risks taken 221 / 359 Risk pays · needless 50 / 666 Safety pays+0.57+0.48 to +0.67 · cautious85.1%188 / 221unstable: fewer than 200 needless and missed risks behind a mean cost +0.07β 1.12 · needless risks lost 0.054 on average (50) · missed risks 0.090 (138)
Claude Fable 5.1highreasoning high · one letter · subscription72.169.5 to 74.952.7%195 / 37091.6%565 / 61769.9%195 / 27938.6%143 / 3708.4%52 / 6178.6%32 / 3701.661.49 to 1.86 · risks taken 227 / 370 Risk pays · needless 52 / 617 Safety pays+0.54+0.43 to +0.67 · cautious85.9%195 / 227unstable: fewer than 200 needless and missed risks behind a mean cost +0.06β 1.11 · needless risks lost 0.076 on average (52) · missed risks 0.114 (143)
Claude Opus 5.5highreasoning high · one letter · subscription70.867.8 to 73.846.3%188 / 40695.3%633 / 66477.0%188 / 24447.5%193 / 4064.7%31 / 6646.2%25 / 4061.731.49 to 2.01 · risks taken 213 / 406 Risk pays · needless 31 / 664 Safety pays+0.80+0.68 to +0.94 · cautious88.3%188 / 213unstable: fewer than 200 needless and missed risks behind a mean cost −0.14β 0.78 · needless risks lost 0.064 on average (31) · missed risks 0.134 (193)
Gemini 4 Argonlowreasoning low · one letter · subscription68.765.6 to 71.949.6%188 / 37987.8%577 / 65764.4%188 / 29244.1%167 / 37912.2%80 / 6576.3%24 / 3791.311.13 to 1.50 · risks taken 212 / 379 Risk pays · needless 80 / 657 Safety pays+0.51+0.42 to +0.60 · cautious88.7%188 / 212unstable: fewer than 200 needless and missed risks behind a mean cost +0.10β 1.14 · needless risks lost 0.092 on average (80) · missed risks 0.140 (167)
Grok 4.7highreasoning high · one letter · subscription67.964.9 to 70.944.6%171 / 38391.1%489 / 53769.0%171 / 24847.8%183 / 3838.9%48 / 5377.6%29 / 3831.401.20 to 1.60 · risks taken 200 / 383 Risk pays · needless 48 / 537 Safety pays+0.64+0.55 to +0.75 · cautious85.5%171 / 200unstable: fewer than 200 needless and missed risks behind a mean cost −0.13β 0.83 · needless risks lost 0.082 on average (48) · missed risks 0.137 (183)
Claude Opus 5.5lowreasoning low · one letter · subscription67.164.5 to 69.541.9%168 / 40192.4%593 / 64269.4%168 / 24251.9%208 / 4017.6%49 / 6426.2%25 / 4011.381.17 to 1.61 · risks taken 193 / 401 Risk pays · needless 49 / 642 Safety pays+0.74+0.63 to +0.87 · cautious87.0%168 / 193unstable: fewer than 200 needless risks behind a mean cost +0.30β 1.51 · needless risks lost 0.108 on average (49) · missed risks 0.115 (208)
Kimi K3.1 Prooffreasoning off · one letter · API66.763.7 to 69.844.7%143 / 32088.7%449 / 50661.4%143 / 23345.0%144 / 32011.3%57 / 50610.3%33 / 3201.331.18 to 1.49 · risks taken 176 / 320 Risk pays · needless 57 / 506 Safety pays+0.54+0.44 to +0.64 · cautious81.2%143 / 176unstable: fewer than 200 needless and missed risks behind a mean cost +0.55β 2.09 · needless risks lost 0.209 on average (57) · missed risks 0.158 (144)
GNUbg weakprogram · CPU · x86-64 build: its noise is shifted, so it overplays gammons60.056.4 to 63.458.1%247 / 42561.8%432 / 69939.4%247 / 62715.3%65 / 42538.2%267 / 69926.6%113 / 4251.321.14 to 1.52 · risks taken 360 / 425 Risk pays · needless 267 / 699 Safety pays−0.36−0.45 to −0.27 · bold68.6%247 / 360unstable: fewer than 200 missed risks behind a mean cost +0.66β 2.38 · needless risks lost 0.132 on average (267) · missed risks 0.091 (65)
Random (computed)reference row, never ranked · computed: every legal play equally likely26.425.9 to 26.923.4%1,054.8 / 4,50629.4%2,178.6 / 7,41811.7%1,054.8 / 8,984.916.9%760.5 / 4,50670.6%5,239.4 / 7,41859.7%2,690.7 / 4,5060.420.39 to 0.44 · risks taken 3,745.5 / 4,506 Risk pays · needless 5,239.4 / 7,418 Safety pays−0.75−0.77 to −0.73 · bold28.2%1,054.8 / 3,745.5+2.51β 2.84 · needless risks lost 0.398 on average (5,239.4) · missed risks 0.231 (760.5)
Safest (computed)reference row, never ranked · computed: always the best safe play50.050.0 to 50.00.0%0 / 4,506100.0%7,418 / 7,418N/A0 / 0100.0%4,506 / 4,5060.0%0 / 7,4180.0%0 / 4,5060.120.11 to 0.14 · risks taken 0 / 4,506 Risk pays · needless 0 / 7,418 Safety pays+3.76+3.75 to +3.76 · cautiousN/A0 / 0N/A: no needless risks to priceneedless risks lost N/A on average (0) · missed risks 0.169 (4,506)

Ability and caution by reasoning setting (UNPAIRED)

Swipe for more columns →

The same model at its lower and higher setting in this league's games: UNPAIRED, the two settings met different positions
ModelAbility d′, 0.001Caution c, 0.001Ability d′, 0.02Caution c, 0.02Median s per decisionOutput tokens per decision (mean)
Claude Opus 5.5low → high · UNPAIRED1.38 → 1.77change +0.39+0.84 → +0.92change +0.091.38 → 1.73change +0.35+0.74 → +0.80change +0.073.92 → 6.35+2.4 s274 → 648
GPT-6 Astralow → medium · UNPAIRED1.85 → 1.78change −0.07+0.65 → +0.70change +0.041.70 → 1.80change +0.10+0.54 → +0.59change +0.053.94 → 4.40+0.5 s98 → 111
GPT-6.1 Sollow → high · UNPAIRED1.68 → 1.67change −0.01+0.69 → +0.69change +0.001.70 → 1.73change +0.03+0.56 → +0.57change +0.014.33 → 13.1+8.8 s114 → 514
Gemini 4 Argonlow → high · UNPAIRED1.42 → 2.04change +0.62+0.61 → +0.91change +0.301.31 → 1.91change +0.60+0.51 → +0.71change +0.213.82 → 27.5+23.7 s27 → 5,417

UNPAIRED: league games are not the same positions, so a change here is not a paired effect of the setting. Output tokens include reasoning tokens where the provider counts them there.

The Phase filter does not apply to this table: all phases.

Analysis pending Shared positions: Analysis pending; no setup has answered them yet.

Risk calls per setup: was each risk the right one, and was safe right when it played safe?

Summary columns first; All columns adds good risks, the error rate by risk-call category, stakes, extra hitting rolls, confidence on the call, ability and caution.

estimate: runs on our own machine (named after it), GPU time, at most; how it was counted on hover or tap · † plays split into groups · Not measured: no value recorded. Method

Setups played different positions; league comparisons are not paired.

Risk

0%0%20%20%40%40%60%60%80%80%100%100%5060708090Risk pays without the right risk (missed + wrong risks): too cautious or wrong play →Needless risks: too bold →Jev 1.13: risk judgement 59.5; missed + wrong 73.0%, needless 7.9%; wrong among them 11.1%Jev 1.13DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release): risk judgement 59.3; missed + wrong 70.9%, needless 10.5%; wrong among them 10.1%DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell): risk judgement 58.8; missed + wrong 67.0%, needless 15.4%; wrong among them 12.1%GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell): risk judgement 57.9; missed + wrong 66.6%, needless 17.5%; wrong among them 15.8%Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)celeris-1-decision: risk judgement 57.8; missed + wrong 76.1%, needless 8.3%; wrong among them 7.1%celeris-1-decisionQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 57.6; missed + wrong 76.5%, needless 8.3%; wrong among them 7.3%Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Cygnet: risk judgement 57.1; missed + wrong 72.7%, needless 13.2%; wrong among them 10.6%CygnetPerplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 56.9; missed + wrong 66.9%, needless 19.3%; wrong among them 14.3%Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)SubQ Snap Preview: risk judgement 56.5; missed + wrong 62.1%, needless 24.9%; wrong among them 26.1%SubQ Snap PreviewGPT-6 Luna (Decisions API): risk judgement 55.7; missed + wrong 78.4%, needless 10.3%; wrong among them 8.8%GPT-6 Luna (Decisions API)Winnow-12B: risk judgement 55.2; missed + wrong 66.3%, needless 23.4%; wrong among them 19.7%Winnow-12BMercury Decide: risk judgement 53.4; missed + wrong 81.3%, needless 11.9%; wrong among them 7.8%Mercury DecideLFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 52.4; missed + wrong 68.5%, needless 26.6%; wrong among them 30.9%LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Perplexity Decider Flash v1.2: risk judgement 49.1; missed + wrong 68.7%, needless 33.1%; wrong among them 37.8%Perplexity Decider Flash v1.2Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell): risk judgement 48.5; missed + wrong 62.7%, needless 40.2%; wrong among them 44.6%Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): risk judgement 45.7; missed + wrong 71.2%, needless 37.3%; wrong among them 31.8%LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)GNUbg weak: risk judgement 67.1; missed + wrong 33.4%, needless 32.4%; wrong among them 68.0%GNUbg weakRandom: risk judgement 24.9; missed + wrong 75.2%, needless 75.0%; wrong among them 74.4%RandomSafest: risk judgement 50.0; missed + wrong 100.0%, needless 0.0%; wrong among them 0.0%Safest

missed risks wrong risks colour: share of wrong risks among the Risk pays decisions without the right risk · squares: reference rows · diagonals: risk judgement

0326598130Jev 1.13decision model · APIneedless risks: 2.10missed risks: 21.41wrong risks: 3.52right risks: 2.12right safety: 2.66No trade-off: 2.6034.4DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)needless risks: 2.94missed risks: 20.07wrong risks: 3.16right risks: 3.52right safety: 3.53No trade-off: 2.3435.6GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsneedless risks: 5.15missed risks: 16.67wrong risks: 2.72right risks: 1.56right safety: 5.81No trade-off: 3.0935.0Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsneedless risks: 6.42missed risks: 16.25wrong risks: 3.85right risks: 2.17right safety: 4.74No trade-off: 3.5136.9celeris-1-decisiondecision model · APIneedless risks: 2.30missed risks: 20.00wrong risks: 2.07right risks: 1.37right safety: 5.50No trade-off: 4.4035.6Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)needless risks: 3.20missed risks: 26.37wrong risks: 2.82right risks: 2.92right safety: 5.75No trade-off: 3.9845.0Cygnetdecision model · our GPUsneedless risks: 4.14missed risks: 19.98wrong risks: 3.19right risks: 2.79right safety: 4.26No trade-off: 4.2038.6Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)decision model · our GPUsneedless risks: 8.32missed risks: 16.99wrong risks: 3.72right risks: 2.11right safety: 3.90No trade-off: 3.7538.8SubQ Snap Previewdecision model · APIneedless risks: 10.40missed risks: 14.35wrong risks: 7.22right risks: 3.21right safety: 5.11No trade-off: 3.5843.9GPT-6 Luna (Decisions API)decision model · APIneedless risks: 3.67missed risks: 25.05wrong risks: 3.45right risks: 2.58right safety: 4.50No trade-off: 4.0743.3Winnow-12Bdecision model · our GPUsneedless risks: 10.05missed risks: 19.93wrong risks: 6.06right risks: 3.45right safety: 5.37No trade-off: 4.8649.7Mercury Decidedecision model · APIneedless risks: 5.59missed risks: 29.17wrong risks: 2.52right risks: 1.67right safety: 4.44No trade-off: 4.5547.9LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)needless risks: 13.29missed risks: 12.21wrong risks: 10.08right risks: 2.73right safety: 4.23No trade-off: 5.1547.7Perplexity Decider Flash v1.2decision model · our GPUsneedless risks: 16.04missed risks: 13.43wrong risks: 13.80right risks: 3.10right safety: 6.17No trade-off: 7.1159.6Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsneedless risks: 21.42missed risks: 12.49wrong risks: 17.29right risks: 3.05right safety: 4.95No trade-off: 6.3565.6LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)needless risks: 18.70missed risks: 14.79wrong risks: 13.15right risks: 2.81right safety: 5.11No trade-off: 6.9061.5GNUbg weakreferenceneedless risks: 8.80missed risks: 1.22wrong risks: 5.99right risks: 2.66right safety: 1.66No trade-off: 1.0721.4Randomreferenceneedless risks: 54.22missed risks: 8.20wrong risks: 38.00right risks: 3.36right safety: 3.88No trade-off: 13.44121.1Safestreferenceneedless risks: 0.00missed risks: 34.95wrong risks: 0.00right risks: 0.00right safety: 8.45No trade-off: 8.2551.6

needless risks missed risks wrong risks right risks right safety No trade-off

0%25%50%75%100%Jev 1.13big needless risks: 29 of 771small stake: found 15.7% (13 / 83)small stake: wrong 13.3% (11 / 83)small stake: missed 71.1% (59 / 83)medium stake: found 18.0% (34 / 189)medium stake: wrong 10.6% (20 / 189)medium stake: missed 71.4% (135 / 189)big stake: found 33.0% (143 / 433)big stake: wrong 6.0% (26 / 433)big stake: missed 61.0% (264 / 433)DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)big needless risks: 38 of 674small stake: found 9.9% (7 / 71)small stake: wrong 16.9% (12 / 71)small stake: missed 73.2% (52 / 71)medium stake: found 17.9% (35 / 196)medium stake: wrong 6.1% (12 / 196)medium stake: missed 76.0% (149 / 196)big stake: found 38.3% (150 / 392)big stake: wrong 5.9% (23 / 392)big stake: missed 55.9% (219 / 392)GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)big needless risks: 75 of 825small stake: found 20.2% (17 / 84)small stake: wrong 15.5% (13 / 84)small stake: missed 64.3% (54 / 84)medium stake: found 28.9% (63 / 218)medium stake: wrong 7.8% (17 / 218)medium stake: missed 63.3% (138 / 218)big stake: found 38.4% (139 / 362)big stake: wrong 6.6% (24 / 362)big stake: missed 55.0% (199 / 362)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)big needless risks: 101 of 827small stake: found 24.1% (21 / 87)small stake: wrong 17.2% (15 / 87)small stake: missed 58.6% (51 / 87)medium stake: found 26.8% (61 / 228)medium stake: wrong 14.5% (33 / 228)medium stake: missed 58.8% (134 / 228)big stake: found 38.9% (162 / 416)big stake: wrong 7.0% (29 / 416)big stake: missed 54.1% (225 / 416)celeris-1-decisionbig needless risks: 38 of 817small stake: found 13.4% (13 / 97)small stake: wrong 6.2% (6 / 97)small stake: missed 80.4% (78 / 97)medium stake: found 14.2% (30 / 212)medium stake: wrong 9.0% (19 / 212)medium stake: missed 76.9% (163 / 212)big stake: found 32.1% (120 / 374)big stake: wrong 3.2% (12 / 374)big stake: missed 64.7% (242 / 374)Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 46 of 732small stake: found 11.8% (8 / 68)small stake: wrong 8.8% (6 / 68)small stake: missed 79.4% (54 / 68)medium stake: found 21.2% (46 / 217)medium stake: wrong 7.8% (17 / 217)medium stake: missed 71.0% (154 / 217)big stake: found 26.3% (124 / 471)big stake: wrong 4.0% (19 / 471)big stake: missed 69.6% (328 / 471)Cygnetbig needless risks: 67 of 789small stake: found 7.1% (5 / 70)small stake: wrong 11.4% (8 / 70)small stake: missed 81.4% (57 / 70)medium stake: found 17.6% (41 / 233)medium stake: wrong 8.2% (19 / 233)medium stake: missed 74.2% (173 / 233)big stake: found 35.6% (159 / 447)big stake: wrong 6.9% (31 / 447)big stake: missed 57.5% (257 / 447)Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 120 of 845small stake: found 13.0% (10 / 77)small stake: wrong 18.2% (14 / 77)small stake: missed 68.8% (53 / 77)medium stake: found 28.0% (63 / 225)medium stake: wrong 10.7% (24 / 225)medium stake: missed 61.3% (138 / 225)big stake: found 39.9% (160 / 401)big stake: wrong 7.2% (29 / 401)big stake: missed 52.9% (212 / 401)SubQ Snap Previewbig needless risks: 156 of 896small stake: found 17.3% (17 / 98)small stake: wrong 24.5% (24 / 98)small stake: missed 58.2% (57 / 98)medium stake: found 32.0% (79 / 247)medium stake: wrong 20.6% (51 / 247)medium stake: missed 47.4% (117 / 247)big stake: found 45.9% (201 / 438)big stake: wrong 11.9% (52 / 438)big stake: missed 42.2% (185 / 438)GPT-6 Luna (Decisions API)big needless risks: 53 of 709small stake: found 9.1% (7 / 77)small stake: wrong 11.7% (9 / 77)small stake: missed 79.2% (61 / 77)medium stake: found 12.9% (30 / 233)medium stake: wrong 7.3% (17 / 233)medium stake: missed 79.8% (186 / 233)big stake: found 28.6% (123 / 430)big stake: wrong 5.8% (25 / 430)big stake: missed 65.6% (282 / 430)Winnow-12Bbig needless risks: 123 of 748small stake: found 11.5% (10 / 87)small stake: wrong 24.1% (21 / 87)small stake: missed 64.4% (56 / 87)medium stake: found 30.6% (77 / 252)medium stake: wrong 16.3% (41 / 252)medium stake: missed 53.2% (134 / 252)big stake: found 39.9% (176 / 441)big stake: wrong 9.1% (40 / 441)big stake: missed 51.0% (225 / 441)Mercury Decidebig needless risks: 51 of 662small stake: found 12.3% (8 / 65)small stake: wrong 3.1% (2 / 65)small stake: missed 84.6% (55 / 65)medium stake: found 14.4% (30 / 208)medium stake: wrong 12.0% (25 / 208)medium stake: missed 73.6% (153 / 208)big stake: found 21.8% (91 / 417)big stake: wrong 4.1% (17 / 417)big stake: missed 74.1% (309 / 417)LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 169 of 882small stake: found 7.5% (7 / 93)small stake: wrong 26.9% (25 / 93)small stake: missed 65.6% (61 / 93)medium stake: found 20.4% (59 / 289)medium stake: wrong 24.9% (72 / 289)medium stake: missed 54.7% (158 / 289)big stake: found 44.8% (182 / 406)big stake: wrong 17.2% (70 / 406)big stake: missed 37.9% (154 / 406)Perplexity Decider Flash v1.2big needless risks: 196 of 776small stake: found 15.3% (13 / 85)small stake: wrong 38.8% (33 / 85)small stake: missed 45.9% (39 / 85)medium stake: found 18.9% (43 / 228)medium stake: wrong 28.5% (65 / 228)medium stake: missed 52.6% (120 / 228)big stake: found 41.9% (167 / 399)big stake: wrong 21.8% (87 / 399)big stake: missed 36.3% (145 / 399)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)big needless risks: 236 of 761small stake: found 17.1% (14 / 82)small stake: wrong 37.8% (31 / 82)small stake: missed 45.1% (37 / 82)medium stake: found 34.9% (80 / 229)medium stake: wrong 33.2% (76 / 229)medium stake: missed 31.9% (73 / 229)big stake: found 42.8% (170 / 397)big stake: wrong 22.9% (91 / 397)big stake: missed 34.3% (136 / 397)LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)big needless risks: 206 of 750small stake: found 11.5% (9 / 78)small stake: wrong 38.5% (30 / 78)small stake: missed 50.0% (39 / 78)medium stake: found 22.3% (50 / 224)medium stake: wrong 21.9% (49 / 224)medium stake: missed 55.8% (125 / 224)big stake: found 35.6% (147 / 413)big stake: wrong 20.1% (83 / 413)big stake: missed 44.3% (183 / 413)GNUbg weakbig needless risks: 142 of 788small stake: found 40.4% (38 / 94)small stake: wrong 39.4% (37 / 94)small stake: missed 20.2% (19 / 94)medium stake: found 54.9% (118 / 215)medium stake: wrong 29.3% (63 / 215)medium stake: missed 15.8% (34 / 215)big stake: found 82.7% (254 / 307)big stake: wrong 13.0% (40 / 307)big stake: missed 4.2% (13 / 307)Randombig needless risks: 329 of 484small stake: found 21.1% (8 / 38)small stake: wrong 57.9% (22 / 38)small stake: missed 21.1% (8 / 38)medium stake: found 17.7% (26 / 147)medium stake: wrong 61.9% (91 / 147)medium stake: missed 20.4% (30 / 147)big stake: found 29.2% (78 / 267)big stake: wrong 52.4% (140 / 267)big stake: missed 18.4% (49 / 267)Safestbig needless risks: 0 of 749small stake: found 0.0% (0 / 78)small stake: wrong 0.0% (0 / 78)small stake: missed 100.0% (78 / 78)medium stake: found 0.0% (0 / 247)medium stake: wrong 0.0% (0 / 247)medium stake: missed 100.0% (247 / 247)big stake: found 0.0% (0 / 460)big stake: wrong 0.0% (0 / 460)big stake: missed 100.0% (460 / 460)

Three bars per setup: small, medium, big stake (Risk pays only). risks found wrong risks missed risks Under each name: needless risks that lost more than 0.08.

-0.500.511.52-1-0.5+0+0.5+1+1.5+2+2.5+3+3.5Ability d′: tells Risk pays from Safety pays →Caution c: + cautious, − boldJev 1.13: ability 1.02, caution +0.90; SDT reference: unstable: fewer than 200 needless risks behind a mean costJev 1.13DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release): ability 0.90, caution +0.80; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell): ability 0.79, caution +0.62; SDT reference: unstable: fewer than 200 needless risks behind a mean costGLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell): ability 0.78, caution +0.54; SDT reference: unstable: fewer than 200 needless risks behind a mean costPerplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)celeris-1-decision: ability 0.84, caution +0.96; SDT reference: unstable: fewer than 200 needless risks behind a mean costceleris-1-decisionQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 0.83, caution +0.96; SDT reference: unstable: fewer than 200 needless risks behind a mean costQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Cygnet: ability 0.73, caution +0.75; SDT reference: unstable: fewer than 200 needless risks behind a mean costCygnetPerplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 0.68, caution +0.53; SDT reference: unstable: fewer than 200 needless risks behind a mean costPerplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)SubQ Snap Preview: ability 0.78, caution +0.29; SDT reference: +0.37SubQ Snap PreviewGPT-6 Luna (Decisions API): ability 0.69, caution +0.91; SDT reference: unstable: fewer than 200 needless risks behind a mean costGPT-6 Luna (Decisions API)Winnow-12B: ability 0.64, caution +0.40; SDT reference: unstable: fewer than 200 needless risks behind a mean costWinnow-12BMercury Decide: ability 0.50, caution +0.92; SDT reference: unstable: fewer than 200 needless risks behind a mean costMercury DecideLFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 0.69, caution +0.28; SDT reference: +0.96LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Perplexity Decider Flash v1.2: ability 0.62, caution +0.13; SDT reference: +0.70Perplexity Decider Flash v1.2Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell): ability 0.64, caution −0.07; SDT reference: +0.62Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): ability 0.36, caution +0.14; SDT reference: +1.38LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)GNUbg weak: ability 1.70, caution −0.39; SDT reference: unstable: fewer than 200 missed risks behind a mean costGNUbg weakRandom: ability 0.19, caution −0.77; SDT reference: unstable: fewer than 200 missed risks behind a mean costRandomSafest: ability −0.01, caution +3.22; SDT reference: N/A: no needless risks to priceSafest

circles: model setups · squares: programs and computed reference rows · short bar: the cost-derived SDT reference c*, drawn only where it is valid

A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity.

0%0%20%20%40%40%60%60%80%80%100%100%5060708090Risk pays without the right risk (missed + wrong risks): too cautious or wrong play →Needless risks: too bold →DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release): risk judgement 59.0; missed + wrong 69.9%, needless 12.1%; wrong among them 10.5%DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)Jev 1.13: risk judgement 58.4; missed + wrong 73.2%, needless 10.0%; wrong among them 12.3%Jev 1.13GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell): risk judgement 57.9; missed + wrong 66.9%, needless 17.3%; wrong among them 13.1%GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)Cygnet: risk judgement 57.3; missed + wrong 71.8%, needless 13.6%; wrong among them 11.9%Cygnetceleris-1-decision: risk judgement 57.3; missed + wrong 75.9%, needless 9.5%; wrong among them 9.0%celeris-1-decisionQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 56.7; missed + wrong 77.2%, needless 9.4%; wrong among them 9.2%Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 56.7; missed + wrong 66.3%, needless 20.3%; wrong among them 15.7%Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell): risk judgement 56.6; missed + wrong 66.9%, needless 19.8%; wrong among them 16.7%Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)SubQ Snap Preview: risk judgement 55.9; missed + wrong 61.6%, needless 26.6%; wrong among them 28.4%SubQ Snap PreviewGPT-6 Luna (Decisions API): risk judgement 55.2; missed + wrong 78.3%, needless 11.3%; wrong among them 9.8%GPT-6 Luna (Decisions API)Winnow-12B: risk judgement 54.5; missed + wrong 66.4%, needless 24.7%; wrong among them 22.0%Winnow-12BMercury Decide: risk judgement 53.1; missed + wrong 81.6%, needless 12.2%; wrong among them 9.4%Mercury DecideLFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 52.4; missed + wrong 67.8%, needless 27.4%; wrong among them 33.8%LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell): risk judgement 48.0; missed + wrong 62.3%, needless 41.6%; wrong among them 46.4%Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)Perplexity Decider Flash v1.2: risk judgement 47.7; missed + wrong 69.4%, needless 35.2%; wrong among them 39.1%Perplexity Decider Flash v1.2LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): risk judgement 44.9; missed + wrong 71.6%, needless 38.5%; wrong among them 32.5%LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)GNUbg weak: risk judgement 64.3; missed + wrong 33.9%, needless 37.4%; wrong among them 73.4%GNUbg weakRandom: risk judgement 23.7; missed + wrong 77.3%, needless 75.3%; wrong among them 75.3%RandomSafest: risk judgement 50.0; missed + wrong 100.0%, needless 0.0%; wrong among them 0.0%Safest

missed risks wrong risks colour: share of wrong risks among the Risk pays decisions without the right risk · squares: reference rows · diagonals: risk judgement

0326598130DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)needless risks: 3.21missed risks: 19.77wrong risks: 3.16right risks: 3.25right safety: 3.83No trade-off: 2.3435.6Jev 1.13decision model · APIneedless risks: 2.36missed risks: 21.14wrong risks: 3.61right risks: 1.77right safety: 2.93No trade-off: 2.6034.4GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsneedless risks: 5.39missed risks: 16.28wrong risks: 2.65right risks: 1.39right safety: 6.20No trade-off: 3.0935.0Cygnetdecision model · our GPUsneedless risks: 4.39missed risks: 19.64wrong risks: 3.24right risks: 2.49right safety: 4.60No trade-off: 4.2038.6celeris-1-decisiondecision model · APIneedless risks: 2.42missed risks: 19.04wrong risks: 2.08right risks: 1.24right safety: 6.47No trade-off: 4.4035.6Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)needless risks: 3.37missed risks: 25.78wrong risks: 3.04right risks: 2.53right safety: 6.34No trade-off: 3.9845.0Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)decision model · our GPUsneedless risks: 8.68missed risks: 16.38wrong risks: 3.80right risks: 1.67right safety: 4.51No trade-off: 3.7538.8Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsneedless risks: 6.93missed risks: 15.82wrong risks: 3.49right risks: 2.02right safety: 5.17No trade-off: 3.5136.9SubQ Snap Previewdecision model · APIneedless risks: 11.35missed risks: 13.80wrong risks: 6.83right risks: 2.65right safety: 5.67No trade-off: 3.5843.9GPT-6 Luna (Decisions API)decision model · APIneedless risks: 4.05missed risks: 24.49wrong risks: 3.47right risks: 2.18right safety: 5.05No trade-off: 4.0743.3Winnow-12Bdecision model · our GPUsneedless risks: 10.78missed risks: 19.41wrong risks: 5.92right risks: 2.87right safety: 5.89No trade-off: 4.8649.7Mercury Decidedecision model · APIneedless risks: 5.60missed risks: 28.68wrong risks: 2.84right risks: 1.34right safety: 4.93No trade-off: 4.5547.9LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)needless risks: 14.80missed risks: 11.60wrong risks: 9.16right risks: 2.14right safety: 4.84No trade-off: 5.1547.7Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsneedless risks: 23.76missed risks: 12.04wrong risks: 15.24right risks: 2.77right safety: 5.40No trade-off: 6.3565.6Perplexity Decider Flash v1.2decision model · our GPUsneedless risks: 17.51missed risks: 13.11wrong risks: 12.97right risks: 2.45right safety: 6.49No trade-off: 7.1159.6LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)needless risks: 20.34missed risks: 14.46wrong risks: 11.97right risks: 2.35right safety: 5.45No trade-off: 6.9061.5GNUbg weakreferenceneedless risks: 9.91missed risks: 1.09wrong risks: 5.28right risks: 2.26right safety: 1.79No trade-off: 1.0721.4Randomreferenceneedless risks: 55.36missed risks: 8.00wrong risks: 37.14right risks: 3.07right safety: 4.08No trade-off: 13.44121.1Safestreferenceneedless risks: 0.00missed risks: 34.38wrong risks: 0.00right risks: 0.00right safety: 9.02No trade-off: 8.2551.6

needless risks missed risks wrong risks right risks right safety No trade-off

0%25%50%75%100%DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)big needless risks: 41 of 745small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 16.8% (33 / 196)medium stake: wrong 7.1% (14 / 196)medium stake: missed 76.0% (149 / 196)big stake: found 36.7% (144 / 392)big stake: wrong 7.4% (29 / 392)big stake: missed 55.9% (219 / 392)Jev 1.13big needless risks: 34 of 854small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 15.9% (30 / 189)medium stake: wrong 12.7% (24 / 189)medium stake: missed 71.4% (135 / 189)big stake: found 31.6% (137 / 433)big stake: wrong 7.4% (32 / 433)big stake: missed 61.0% (264 / 433)GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)big needless risks: 79 of 909small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 24.8% (54 / 218)medium stake: wrong 11.9% (26 / 218)medium stake: missed 63.3% (138 / 218)big stake: found 38.1% (138 / 362)big stake: wrong 6.9% (25 / 362)big stake: missed 55.0% (199 / 362)Cygnetbig needless risks: 73 of 859small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 16.7% (39 / 233)medium stake: wrong 9.0% (21 / 233)medium stake: missed 74.2% (173 / 233)big stake: found 34.2% (153 / 447)big stake: wrong 8.3% (37 / 447)big stake: missed 57.5% (257 / 447)celeris-1-decisionbig needless risks: 41 of 914small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 11.3% (24 / 212)medium stake: wrong 11.8% (25 / 212)medium stake: missed 76.9% (163 / 212)big stake: found 31.3% (117 / 374)big stake: wrong 4.0% (15 / 374)big stake: missed 64.7% (242 / 374)Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 51 of 800small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 18.0% (39 / 217)medium stake: wrong 11.1% (24 / 217)medium stake: missed 71.0% (154 / 217)big stake: found 25.1% (118 / 471)big stake: wrong 5.3% (25 / 471)big stake: missed 69.6% (328 / 471)Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 126 of 922small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 25.3% (57 / 225)medium stake: wrong 13.3% (30 / 225)medium stake: missed 61.3% (138 / 225)big stake: found 38.4% (154 / 401)big stake: wrong 8.7% (35 / 401)big stake: missed 52.9% (212 / 401)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)big needless risks: 109 of 914small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 24.6% (56 / 228)medium stake: wrong 16.7% (38 / 228)medium stake: missed 58.8% (134 / 228)big stake: found 37.7% (157 / 416)big stake: wrong 8.2% (34 / 416)big stake: missed 54.1% (225 / 416)SubQ Snap Previewbig needless risks: 172 of 994small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 29.6% (73 / 247)medium stake: wrong 23.1% (57 / 247)medium stake: missed 47.4% (117 / 247)big stake: found 43.4% (190 / 438)big stake: wrong 14.4% (63 / 438)big stake: missed 42.2% (185 / 438)GPT-6 Luna (Decisions API)big needless risks: 59 of 786small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 11.6% (27 / 233)medium stake: wrong 8.6% (20 / 233)medium stake: missed 79.8% (186 / 233)big stake: found 27.2% (117 / 430)big stake: wrong 7.2% (31 / 430)big stake: missed 65.6% (282 / 430)Winnow-12Bbig needless risks: 132 of 835small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 27.0% (68 / 252)medium stake: wrong 19.8% (50 / 252)medium stake: missed 53.2% (134 / 252)big stake: found 37.4% (165 / 441)big stake: wrong 11.6% (51 / 441)big stake: missed 51.0% (225 / 441)Mercury Decidebig needless risks: 51 of 727small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 13.5% (28 / 208)medium stake: wrong 13.0% (27 / 208)medium stake: missed 73.6% (153 / 208)big stake: found 20.9% (87 / 417)big stake: wrong 5.0% (21 / 417)big stake: missed 74.1% (309 / 417)LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 186 of 975small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 18.3% (53 / 289)medium stake: wrong 27.0% (78 / 289)medium stake: missed 54.7% (158 / 289)big stake: found 42.1% (171 / 406)big stake: wrong 20.0% (81 / 406)big stake: missed 37.9% (154 / 406)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)big needless risks: 261 of 843small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 31.4% (72 / 229)medium stake: wrong 36.7% (84 / 229)medium stake: missed 31.9% (73 / 229)big stake: found 41.3% (164 / 397)big stake: wrong 24.4% (97 / 397)big stake: missed 34.3% (136 / 397)Perplexity Decider Flash v1.2big needless risks: 219 of 861small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 14.9% (34 / 228)medium stake: wrong 32.5% (74 / 228)medium stake: missed 52.6% (120 / 228)big stake: found 39.6% (158 / 399)big stake: wrong 24.1% (96 / 399)big stake: missed 36.3% (145 / 399)LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)big needless risks: 226 of 828small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 19.6% (44 / 224)medium stake: wrong 24.6% (55 / 224)medium stake: missed 55.8% (125 / 224)big stake: found 33.2% (137 / 413)big stake: wrong 22.5% (93 / 413)big stake: missed 44.3% (183 / 413)GNUbg weakbig needless risks: 159 of 882small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 44.7% (96 / 215)medium stake: wrong 39.5% (85 / 215)medium stake: missed 15.8% (34 / 215)big stake: found 81.1% (249 / 307)big stake: wrong 14.7% (45 / 307)big stake: missed 4.2% (13 / 307)Randombig needless risks: 346 of 522small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 12.9% (19 / 147)medium stake: wrong 66.7% (98 / 147)medium stake: missed 20.4% (30 / 147)big stake: found 28.1% (75 / 267)big stake: wrong 53.6% (143 / 267)big stake: missed 18.4% (49 / 267)Safestbig needless risks: 0 of 827small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 0.0% (0 / 247)medium stake: wrong 0.0% (0 / 247)medium stake: missed 100.0% (247 / 247)big stake: found 0.0% (0 / 460)big stake: wrong 0.0% (0 / 460)big stake: missed 100.0% (460 / 460)

Three bars per setup: small, medium, big stake (Risk pays only). risks found wrong risks missed risks Under each name: needless risks that lost more than 0.08.

-0.500.511.52-1-0.5+0+0.5+1+1.5+2+2.5+3+3.5Ability d′: tells Risk pays from Safety pays →Caution c: + cautious, − boldDeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release): ability 0.85, caution +0.74; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)Jev 1.13: ability 0.92, caution +0.82; SDT reference: unstable: fewer than 200 needless risks behind a mean costJev 1.13GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell): ability 0.74, caution +0.57; SDT reference: unstable: fewer than 200 needless risks behind a mean costGLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)Cygnet: ability 0.76, caution +0.72; SDT reference: unstable: fewer than 200 needless risks behind a mean costCygnetceleris-1-decision: ability 0.81, caution +0.90; SDT reference: unstable: fewer than 200 needless risks behind a mean costceleris-1-decisionQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 0.79, caution +0.92; SDT reference: unstable: fewer than 200 needless risks behind a mean costQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 0.68, caution +0.49; SDT reference: unstable: fewer than 200 needless risks behind a mean costPerplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell): ability 0.70, caution +0.50; SDT reference: unstable: fewer than 200 needless risks behind a mean costPerplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)SubQ Snap Preview: ability 0.77, caution +0.24; SDT reference: +0.40SubQ Snap PreviewGPT-6 Luna (Decisions API): ability 0.67, caution +0.87; SDT reference: unstable: fewer than 200 needless risks behind a mean costGPT-6 Luna (Decisions API)Winnow-12B: ability 0.64, caution +0.36; SDT reference: +0.24Winnow-12BMercury Decide: ability 0.52, caution +0.90; SDT reference: unstable: fewer than 200 needless risks behind a mean costMercury DecideLFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 0.73, caution +0.24; SDT reference: +1.01LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell): ability 0.64, caution −0.11; SDT reference: +0.72Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)Perplexity Decider Flash v1.2: ability 0.57, caution +0.09; SDT reference: +0.82Perplexity Decider Flash v1.2LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): ability 0.33, caution +0.13; SDT reference: +1.71LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)GNUbg weak: ability 1.66, caution −0.51; SDT reference: unstable: fewer than 200 missed risks behind a mean costGNUbg weakRandom: ability 0.19, caution −0.78; SDT reference: unstable: fewer than 200 missed risks behind a mean costRandomSafest: ability 0.04, caution +3.21; SDT reference: N/A: no needless risks to priceSafest

circles: model setups · squares: programs and computed reference rows · short bar: the cost-derived SDT reference c*, drawn only where it is valid

A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity.

0%0%20%20%40%40%60%60%80%80%100%100%5060708090Risk pays without the right risk (missed + wrong risks): too cautious or wrong play →Needless risks: too bold →DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release): risk judgement 69.0; missed + wrong 55.0%, needless 7.0%; wrong among them 10.9%DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)DeepSeek-V4.1-Flash (low · API · 1017): risk judgement 68.5; missed + wrong 57.5%, needless 5.5%; wrong among them 9.4%DeepSeek-V4.1-Flash (low · API · 1017)Claude Sonnet 5.5 (high): risk judgement 68.2; missed + wrong 58.2%, needless 5.4%; wrong among them 4.7%Claude Sonnet 5.5 (high)Claude Fable 5.1 (low): risk judgement 67.8; missed + wrong 59.0%, needless 5.4%; wrong among them 8.8%Claude Fable 5.1 (low)Grok 4.7 (low): risk judgement 67.7; missed + wrong 57.7%, needless 7.0%; wrong among them 13.4%Grok 4.7 (low)Gemini 3.8 Flash (high): risk judgement 66.9; missed + wrong 63.9%, needless 2.4%; wrong among them 3.6%Gemini 3.8 Flash (high)Gemini 3.8 Flash (low): risk judgement 65.1; missed + wrong 66.2%, needless 3.5%; wrong among them 4.5%Gemini 3.8 Flash (low)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 64.5; missed + wrong 68.1%, needless 2.9%; wrong among them 4.8%Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)d1-ultra: risk judgement 64.0; missed + wrong 58.0%, needless 14.0%; wrong among them 14.5%d1-ultraClaude Sonnet 5.5 (low): risk judgement 63.5; missed + wrong 69.1%, needless 3.8%; wrong among them 5.3%Claude Sonnet 5.5 (low)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 62.6; missed + wrong 66.8%, needless 7.9%; wrong among them 12.6%Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 62.0; missed + wrong 66.2%, needless 9.8%; wrong among them 11.8%Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)GLM-5.3 (low): risk judgement 61.7; missed + wrong 69.3%, needless 7.3%; wrong among them 8.9%GLM-5.3 (low)GLiDE (API): risk judgement 60.4; missed + wrong 63.1%, needless 16.0%; wrong among them 19.2%GLiDE (API)Qwen3.8 Flash (off): risk judgement 59.5; missed + wrong 65.6%, needless 15.4%; wrong among them 19.7%Qwen3.8 Flash (off)Jev 1.14: risk judgement 59.3; missed + wrong 70.9%, needless 10.6%; wrong among them 10.5%Jev 1.14DeepSeek-V4.1-Flash (off · API · 1017): risk judgement 58.5; missed + wrong 74.6%, needless 8.3%; wrong among them 8.5%DeepSeek-V4.1-Flash (off · API · 1017)Clef: risk judgement 57.5; missed + wrong 60.5%, needless 24.4%; wrong among them 25.3%ClefGNUbg weak: risk judgement 62.3; missed + wrong 37.3%, needless 38.1%; wrong among them 61.3%GNUbg weakRandom (computed): risk judgement 27.0; missed + wrong 74.9%, needless 71.2%; wrong among them 77.7%Random (computed)Safest (computed): risk judgement 50.0; missed + wrong 100.0%, needless 0.0%; wrong among them 0.0%Safest (computed)

missed risks wrong risks colour: share of wrong risks among the Risk pays decisions without the right risk · squares: reference rows · diagonals: risk judgement

0306090120DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)reasoning low · letter probabilities · APIneedless risks: 2.27missed risks: 8.56wrong risks: 1.93right risks: 2.41right safety: 2.89No trade-off: 0.9319.0DeepSeek-V4.1-Flash (low · API · 1017)reasoning low · letter probabilities · APIneedless risks: 1.48missed risks: 9.93wrong risks: 1.57right risks: 1.68right safety: 1.63No trade-off: 0.8617.1Claude Sonnet 5.5 (high)reasoning high · one letter · subscriptionneedless risks: 1.12missed risks: 10.76wrong risks: 0.64right risks: 1.70right safety: 1.58No trade-off: 0.9216.7Claude Fable 5.1 (low)needless risks: 1.59missed risks: 8.82wrong risks: 1.49right risks: 1.74right safety: 2.68No trade-off: 1.6117.9Grok 4.7 (low)needless risks: 1.99missed risks: 10.40wrong risks: 2.28right risks: 1.90right safety: 2.04No trade-off: 0.5119.1Gemini 3.8 Flash (high)reasoning high · one letter · subscriptionneedless risks: 0.42missed risks: 10.77wrong risks: 0.59right risks: 1.16right safety: 1.50No trade-off: 0.5014.9Gemini 3.8 Flash (low)reasoning low · one letter · subscriptionneedless risks: 0.80missed risks: 9.90wrong risks: 0.74right risks: 0.79right safety: 2.82No trade-off: 1.2916.3Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning on · letter probabilities · our GPUsneedless risks: 0.54missed risks: 12.79wrong risks: 1.21right risks: 0.96right safety: 3.00No trade-off: 1.0819.6d1-ultraneedless risks: 3.66missed risks: 13.25wrong risks: 3.02right risks: 2.37right safety: 4.52No trade-off: 3.5130.3Claude Sonnet 5.5 (low)reasoning low · one letter · subscriptionneedless risks: 1.22missed risks: 12.96wrong risks: 0.96right risks: 1.07right safety: 2.67No trade-off: 1.2620.1Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning off · letter probabilities · our GPUsneedless risks: 2.06missed risks: 11.12wrong risks: 2.12right risks: 1.55right safety: 3.07No trade-off: 2.5722.5Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning low · letter probabilities · our GPUsneedless risks: 2.89missed risks: 14.82wrong risks: 2.55right risks: 1.94right safety: 3.41No trade-off: 3.4329.0GLM-5.3 (low)reasoning low · letter probabilities · APIneedless risks: 1.87missed risks: 17.40wrong risks: 2.14right risks: 1.78right safety: 3.99No trade-off: 3.7030.9GLiDE (API)needless risks: 7.18missed risks: 10.97wrong risks: 4.79right risks: 2.03right safety: 4.97No trade-off: 3.2433.2Qwen3.8 Flash (off)reasoning off · letter probabilities · APIneedless risks: 4.64missed risks: 14.39wrong risks: 3.96right risks: 3.46right safety: 2.96No trade-off: 4.1733.6Jev 1.14needless risks: 2.84missed risks: 16.00wrong risks: 2.91right risks: 1.72right safety: 4.37No trade-off: 3.8931.7DeepSeek-V4.1-Flash (off · API · 1017)reasoning off · letter probabilities · APIneedless risks: 2.49missed risks: 21.43wrong risks: 2.50right risks: 1.61right safety: 3.75No trade-off: 3.4735.2Clefneedless risks: 8.85missed risks: 11.85wrong risks: 5.48right risks: 2.26right safety: 4.18No trade-off: 3.9236.5GNUbg weakreferenceneedless risks: 12.05missed risks: 2.19wrong risks: 6.03right risks: 2.93right safety: 1.97No trade-off: 1.8027.0Random (computed)reference, computedneedless risks: 57.31missed risks: 5.40wrong risks: 40.03right risks: 3.23right safety: 3.91No trade-off: 8.04117.9Safest (computed)reference, computedneedless risks: 0.00missed risks: 25.23wrong risks: 0.00right risks: 0.00right safety: 0.00No trade-off: 0.0025.2

needless risks missed risks wrong risks right risks right safety No trade-off

0%25%50%75%100%DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)big needless risks: 33 of 958small stake: found 11.7% (13 / 111)small stake: wrong 12.6% (14 / 111)small stake: missed 75.7% (84 / 111)medium stake: found 29.9% (70 / 234)medium stake: wrong 7.3% (17 / 234)medium stake: missed 62.8% (147 / 234)big stake: found 62.8% (255 / 406)big stake: wrong 3.4% (14 / 406)big stake: missed 33.7% (137 / 406)DeepSeek-V4.1-Flash (low · API · 1017)big needless risks: 26 of 923small stake: found 13.9% (14 / 101)small stake: wrong 6.9% (7 / 101)small stake: missed 79.2% (80 / 101)medium stake: found 32.9% (81 / 246)medium stake: wrong 6.9% (17 / 246)medium stake: missed 60.2% (148 / 246)big stake: found 55.4% (226 / 408)big stake: wrong 4.2% (17 / 408)big stake: missed 40.4% (165 / 408)Claude Sonnet 5.5 (high)big needless risks: 17 of 889small stake: found 12.3% (9 / 73)small stake: wrong 1.4% (1 / 73)small stake: missed 86.3% (63 / 73)medium stake: found 27.1% (69 / 255)medium stake: wrong 5.5% (14 / 255)medium stake: missed 67.5% (172 / 255)big stake: found 56.2% (230 / 409)big stake: wrong 1.2% (5 / 409)big stake: missed 42.5% (174 / 409)Claude Fable 5.1 (low)big needless risks: 28 of 826small stake: found 12.1% (11 / 91)small stake: wrong 6.6% (6 / 91)small stake: missed 81.3% (74 / 91)medium stake: found 26.9% (61 / 227)medium stake: wrong 8.4% (19 / 227)medium stake: missed 64.8% (147 / 227)big stake: found 58.2% (198 / 340)big stake: wrong 2.6% (9 / 340)big stake: missed 39.1% (133 / 340)Grok 4.7 (low)big needless risks: 31 of 916small stake: found 21.7% (18 / 83)small stake: wrong 8.4% (7 / 83)small stake: missed 69.9% (58 / 83)medium stake: found 34.1% (76 / 223)medium stake: wrong 11.7% (26 / 223)medium stake: missed 54.3% (121 / 223)big stake: found 51.4% (201 / 391)big stake: wrong 5.4% (21 / 391)big stake: missed 43.2% (169 / 391)Gemini 3.8 Flash (high)big needless risks: 4 of 879small stake: found 11.6% (14 / 121)small stake: wrong 3.3% (4 / 121)small stake: missed 85.1% (103 / 121)medium stake: found 23.3% (57 / 245)medium stake: wrong 3.3% (8 / 245)medium stake: missed 73.5% (180 / 245)big stake: found 52.9% (193 / 365)big stake: wrong 1.4% (5 / 365)big stake: missed 45.8% (167 / 365)Gemini 3.8 Flash (low)big needless risks: 15 of 909small stake: found 5.4% (5 / 93)small stake: wrong 3.2% (3 / 93)small stake: missed 91.4% (85 / 93)medium stake: found 24.7% (58 / 235)medium stake: wrong 4.3% (10 / 235)medium stake: missed 71.1% (167 / 235)big stake: found 49.5% (151 / 305)big stake: wrong 2.0% (6 / 305)big stake: missed 48.5% (148 / 305)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 11 of 904small stake: found 4.6% (4 / 87)small stake: wrong 5.7% (5 / 87)small stake: missed 89.7% (78 / 87)medium stake: found 23.1% (55 / 238)medium stake: wrong 2.5% (6 / 238)medium stake: missed 74.4% (177 / 238)big stake: found 43.7% (164 / 375)big stake: wrong 3.2% (12 / 375)big stake: missed 53.1% (199 / 375)d1-ultrabig needless risks: 60 of 816small stake: found 25.9% (21 / 81)small stake: wrong 8.6% (7 / 81)small stake: missed 65.4% (53 / 81)medium stake: found 37.3% (88 / 236)medium stake: wrong 11.9% (28 / 236)medium stake: missed 50.8% (120 / 236)big stake: found 48.5% (180 / 371)big stake: wrong 6.2% (23 / 371)big stake: missed 45.3% (168 / 371)Claude Sonnet 5.5 (low)big needless risks: 25 of 918small stake: found 10.3% (9 / 87)small stake: wrong 4.6% (4 / 87)small stake: missed 85.1% (74 / 87)medium stake: found 21.9% (50 / 228)medium stake: wrong 4.8% (11 / 228)medium stake: missed 73.2% (167 / 228)big stake: found 41.9% (145 / 346)big stake: wrong 2.6% (9 / 346)big stake: missed 55.5% (192 / 346)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 39 of 795small stake: found 9.9% (7 / 71)small stake: wrong 12.7% (9 / 71)small stake: missed 77.5% (55 / 71)medium stake: found 24.4% (51 / 209)medium stake: wrong 10.0% (21 / 209)medium stake: missed 65.6% (137 / 209)big stake: found 44.7% (136 / 304)big stake: wrong 6.2% (19 / 304)big stake: missed 49.0% (149 / 304)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 49 of 815small stake: found 16.5% (14 / 85)small stake: wrong 17.6% (15 / 85)small stake: missed 65.9% (56 / 85)medium stake: found 24.8% (54 / 218)medium stake: wrong 10.1% (22 / 218)medium stake: missed 65.1% (142 / 218)big stake: found 42.2% (174 / 412)big stake: wrong 4.6% (19 / 412)big stake: missed 53.2% (219 / 412)GLM-5.3 (low)big needless risks: 35 of 783small stake: found 9.0% (8 / 89)small stake: wrong 11.2% (10 / 89)small stake: missed 79.8% (71 / 89)medium stake: found 28.6% (56 / 196)medium stake: wrong 6.6% (13 / 196)medium stake: missed 64.8% (127 / 196)big stake: found 36.8% (141 / 383)big stake: wrong 4.7% (18 / 383)big stake: missed 58.5% (224 / 383)GLiDE (API)big needless risks: 107 of 900small stake: found 17.1% (14 / 82)small stake: wrong 17.1% (14 / 82)small stake: missed 65.9% (54 / 82)medium stake: found 29.6% (66 / 223)medium stake: wrong 10.8% (24 / 223)medium stake: missed 59.6% (133 / 223)big stake: found 45.7% (169 / 370)big stake: wrong 11.9% (44 / 370)big stake: missed 42.4% (157 / 370)Qwen3.8 Flash (off)big needless risks: 69 of 725small stake: found 13.6% (11 / 81)small stake: wrong 17.3% (14 / 81)small stake: missed 69.1% (56 / 81)medium stake: found 28.2% (66 / 234)medium stake: wrong 17.9% (42 / 234)medium stake: missed 53.8% (126 / 234)big stake: found 42.5% (166 / 391)big stake: wrong 9.0% (35 / 391)big stake: missed 48.6% (190 / 391)Jev 1.14big needless risks: 45 of 735small stake: found 12.3% (9 / 73)small stake: wrong 8.2% (6 / 73)small stake: missed 79.5% (58 / 73)medium stake: found 27.9% (57 / 204)medium stake: wrong 6.9% (14 / 204)medium stake: missed 65.2% (133 / 204)big stake: found 33.2% (118 / 355)big stake: wrong 7.6% (27 / 355)big stake: missed 59.2% (210 / 355)DeepSeek-V4.1-Flash (off · API · 1017)big needless risks: 38 of 745small stake: found 7.7% (6 / 78)small stake: wrong 6.4% (5 / 78)small stake: missed 85.9% (67 / 78)medium stake: found 16.8% (37 / 220)medium stake: wrong 9.1% (20 / 220)medium stake: missed 74.1% (163 / 220)big stake: found 33.3% (138 / 415)big stake: wrong 4.8% (20 / 415)big stake: missed 61.9% (257 / 415)Clefbig needless risks: 123 of 798small stake: found 20.0% (15 / 75)small stake: wrong 16.0% (12 / 75)small stake: missed 64.0% (48 / 75)medium stake: found 31.9% (66 / 207)medium stake: wrong 21.3% (44 / 207)medium stake: missed 46.9% (97 / 207)big stake: found 47.4% (182 / 384)big stake: wrong 12.0% (46 / 384)big stake: missed 40.6% (156 / 384)GNUbg weakbig needless risks: 211 of 902small stake: found 26.2% (32 / 122)small stake: wrong 40.2% (49 / 122)small stake: missed 33.6% (41 / 122)medium stake: found 49.4% (124 / 251)medium stake: wrong 34.7% (87 / 251)medium stake: missed 15.9% (40 / 251)big stake: found 82.4% (327 / 397)big stake: wrong 10.1% (40 / 397)big stake: missed 7.6% (30 / 397)Random (computed)big needless risks: 10,184.8 of 16,136small stake: found 12.8% (214.8 / 1,684)small stake: wrong 66.7% (1,123.1 / 1,684)small stake: missed 20.6% (346.1 / 1,684)medium stake: found 19.3% (836.6 / 4,329)medium stake: wrong 63.3% (2,741.1 / 4,329)medium stake: missed 17.4% (751.3 / 4,329)big stake: found 31.6% (2,248.9 / 7,127)big stake: wrong 53.1% (3,782.5 / 7,127)big stake: missed 15.4% (1,095.6 / 7,127)Safest (computed)big needless risks: 0 of 16,136small stake: found 0.0% (0 / 1,684)small stake: wrong 0.0% (0 / 1,684)small stake: missed 100.0% (1,684 / 1,684)medium stake: found 0.0% (0 / 4,329)medium stake: wrong 0.0% (0 / 4,329)medium stake: missed 100.0% (4,329 / 4,329)big stake: found 0.0% (0 / 7,127)big stake: wrong 0.0% (0 / 7,127)big stake: missed 100.0% (7,127 / 7,127)

Three bars per setup: small, medium, big stake (Risk pays only). risks found wrong risks missed risks Under each name: needless risks that lost more than 0.08.

-0.500.511.52-1-0.5+0+0.5+1+1.5+2+2.5+3+3.5+4Ability d′: tells Risk pays from Safety pays →Caution c: + cautious, − boldDeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release): ability 1.50, caution +0.72; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)DeepSeek-V4.1-Flash (low · API · 1017): ability 1.54, caution +0.82; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (low · API · 1017)Claude Sonnet 5.5 (high): ability 1.46, caution +0.87; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Sonnet 5.5 (high)Claude Fable 5.1 (low): ability 1.50, caution +0.85; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Fable 5.1 (low)Grok 4.7 (low): ability 1.48, caution +0.74; SDT reference: unstable: fewer than 200 needless risks behind a mean costGrok 4.7 (low)Gemini 3.8 Flash (high): ability 1.68, caution +1.13; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemini 3.8 Flash (high)Gemini 3.8 Flash (low): ability 1.47, caution +1.07; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemini 3.8 Flash (low)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 1.51, caution +1.14; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)d1-ultra: ability 1.09, caution +0.53; SDT reference: unstable: fewer than 200 needless risks behind a mean costd1-ultraClaude Sonnet 5.5 (low): ability 1.37, caution +1.08; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Sonnet 5.5 (low)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 1.20, caution +0.81; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 1.08, caution +0.75; SDT reference: unstable: fewer than 200 needless risks behind a mean costQwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)GLM-5.3 (low): ability 1.12, caution +0.89; SDT reference: unstable: fewer than 200 needless risks behind a mean costGLM-5.3 (low)GLiDE (API): ability 0.97, caution +0.51; SDT reference: unstable: fewer than 200 needless risks behind a mean costGLiDE (API)Qwen3.8 Flash (off): ability 0.95, caution +0.54; SDT reference: unstable: fewer than 200 needless risks behind a mean costQwen3.8 Flash (off)Jev 1.14: ability 0.90, caution +0.79; SDT reference: unstable: fewer than 200 needless risks behind a mean costJev 1.14DeepSeek-V4.1-Flash (off · API · 1017): ability 0.90, caution +0.93; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (off · API · 1017)Clef: ability 0.81, caution +0.29; SDT reference: unstable: fewer than 200 needless risks behind a mean costClefGNUbg weak: ability 1.36, caution −0.38; SDT reference: unstable: fewer than 200 missed risks behind a mean costGNUbg weakRandom (computed): ability 0.41, caution −0.76; SDT reference: +2.23Random (computed)Safest (computed): ability 0.05, caution +3.98; SDT reference: N/A: no needless risks to priceSafest (computed)

circles: model setups · squares: programs and computed reference rows · short bar: the cost-derived SDT reference c*, drawn only where it is valid

A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity.

0%0%20%20%40%40%60%60%80%80%100%100%5060708090Risk pays without the right risk (missed + wrong risks): too cautious or wrong play →Needless risks: too bold →DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release): risk judgement 70.2; missed + wrong 50.8%, needless 8.8%; wrong among them 12.6%DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)Claude Sonnet 5.5 (high): risk judgement 69.0; missed + wrong 56.0%, needless 6.0%; wrong among them 7.0%Claude Sonnet 5.5 (high)DeepSeek-V4.1-Flash (low · API · 1017): risk judgement 68.9; missed + wrong 55.2%, needless 7.0%; wrong among them 13.3%DeepSeek-V4.1-Flash (low · API · 1017)Claude Fable 5.1 (low): risk judgement 68.5; missed + wrong 56.3%, needless 6.8%; wrong among them 12.2%Claude Fable 5.1 (low)Gemini 3.8 Flash (high): risk judgement 67.6; missed + wrong 60.8%, needless 3.9%; wrong among them 6.5%Gemini 3.8 Flash (high)Grok 4.7 (low): risk judgement 67.0; missed + wrong 57.2%, needless 8.9%; wrong among them 17.4%Grok 4.7 (low)Gemini 3.8 Flash (low): risk judgement 66.8; missed + wrong 62.4%, needless 4.0%; wrong among them 6.5%Gemini 3.8 Flash (low)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 65.6; missed + wrong 65.3%, needless 3.5%; wrong among them 6.0%Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Claude Sonnet 5.5 (low): risk judgement 64.2; missed + wrong 66.9%, needless 4.8%; wrong among them 6.5%Claude Sonnet 5.5 (low)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 63.2; missed + wrong 64.5%, needless 9.1%; wrong among them 13.6%Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)d1-ultra: risk judgement 63.0; missed + wrong 58.2%, needless 15.8%; wrong among them 18.4%d1-ultraGLM-5.3 (low): risk judgement 61.6; missed + wrong 68.2%, needless 8.6%; wrong among them 11.1%GLM-5.3 (low)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risk judgement 61.2; missed + wrong 65.6%, needless 12.1%; wrong among them 12.6%Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)GLiDE (API): risk judgement 60.5; missed + wrong 61.6%, needless 17.5%; wrong among them 20.5%GLiDE (API)Qwen3.8 Flash (off): risk judgement 59.3; missed + wrong 64.5%, needless 17.0%; wrong among them 21.6%Qwen3.8 Flash (off)Jev 1.14: risk judgement 59.2; missed + wrong 70.1%, needless 11.5%; wrong among them 12.5%Jev 1.14DeepSeek-V4.1-Flash (off · API · 1017): risk judgement 58.8; missed + wrong 73.5%, needless 8.9%; wrong among them 10.1%DeepSeek-V4.1-Flash (off · API · 1017)Clef: risk judgement 57.6; missed + wrong 59.4%, needless 25.4%; wrong among them 27.9%ClefGNUbg weak: risk judgement 61.7; missed + wrong 35.2%, needless 41.5%; wrong among them 69.3%GNUbg weakRandom (computed): risk judgement 25.9; missed + wrong 76.3%, needless 71.9%; wrong among them 78.9%Random (computed)Safest (computed): risk judgement 50.0; missed + wrong 100.0%, needless 0.0%; wrong among them 0.0%Safest (computed)

missed risks wrong risks colour: share of wrong risks among the Risk pays decisions without the right risk · squares: reference rows · diagonals: risk judgement

0306090120DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)reasoning low · letter probabilities · APIneedless risks: 2.67missed risks: 8.23wrong risks: 1.70right risks: 2.24right safety: 3.21No trade-off: 0.9319.0Claude Sonnet 5.5 (high)reasoning high · one letter · subscriptionneedless risks: 1.13missed risks: 10.53wrong risks: 0.77right risks: 1.56right safety: 1.81No trade-off: 0.9216.7DeepSeek-V4.1-Flash (low · API · 1017)reasoning low · letter probabilities · APIneedless risks: 1.58missed risks: 9.49wrong risks: 1.73right risks: 1.42right safety: 2.07No trade-off: 0.8617.1Claude Fable 5.1 (low)needless risks: 1.89missed risks: 8.37wrong risks: 1.40right risks: 1.53right safety: 3.13No trade-off: 1.6117.9Gemini 3.8 Flash (high)reasoning high · one letter · subscriptionneedless risks: 0.44missed risks: 10.25wrong risks: 0.70right risks: 1.02right safety: 2.02No trade-off: 0.5014.9Grok 4.7 (low)needless risks: 2.15missed risks: 10.12wrong risks: 2.37right risks: 1.66right safety: 2.32No trade-off: 0.5119.1Gemini 3.8 Flash (low)reasoning low · one letter · subscriptionneedless risks: 0.91missed risks: 9.52wrong risks: 0.73right risks: 0.69right safety: 3.20No trade-off: 1.2916.3Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning on · letter probabilities · our GPUsneedless risks: 0.78missed risks: 12.28wrong risks: 1.02right risks: 0.91right safety: 3.51No trade-off: 1.0819.6Claude Sonnet 5.5 (low)reasoning low · one letter · subscriptionneedless risks: 1.28missed risks: 12.64wrong risks: 1.00right risks: 0.98right safety: 2.99No trade-off: 1.2620.1Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning off · letter probabilities · our GPUsneedless risks: 2.27missed risks: 10.68wrong risks: 2.02right risks: 1.43right safety: 3.51No trade-off: 2.5722.5d1-ultraneedless risks: 4.19missed risks: 12.84wrong risks: 3.03right risks: 1.84right safety: 4.93No trade-off: 3.5130.3GLM-5.3 (low)reasoning low · letter probabilities · APIneedless risks: 2.12missed risks: 17.02wrong risks: 2.39right risks: 1.28right safety: 4.36No trade-off: 3.7030.9Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning low · letter probabilities · our GPUsneedless risks: 3.39missed risks: 14.46wrong risks: 2.43right risks: 1.56right safety: 3.76No trade-off: 3.4329.0GLiDE (API)needless risks: 7.56missed risks: 10.52wrong risks: 4.52right risks: 1.91right safety: 5.42No trade-off: 3.2433.2Qwen3.8 Flash (off)reasoning off · letter probabilities · APIneedless risks: 4.98missed risks: 14.05wrong risks: 4.03right risks: 3.05right safety: 3.30No trade-off: 4.1733.6Jev 1.14needless risks: 2.93missed risks: 15.64wrong risks: 3.02right risks: 1.53right safety: 4.73No trade-off: 3.8931.7DeepSeek-V4.1-Flash (off · API · 1017)reasoning off · letter probabilities · APIneedless risks: 2.60missed risks: 21.03wrong risks: 2.59right risks: 1.42right safety: 4.15No trade-off: 3.4735.2Clefneedless risks: 9.44missed risks: 11.41wrong risks: 5.11right risks: 2.04right safety: 4.62No trade-off: 3.9236.5GNUbg weakreferenceneedless risks: 12.91missed risks: 2.00wrong risks: 5.60right risks: 2.50right safety: 2.15No trade-off: 1.8027.0Random (computed)reference, computedneedless risks: 61.87missed risks: 5.15wrong risks: 36.08right risks: 2.62right safety: 4.17No trade-off: 8.04117.9Safest (computed)reference, computedneedless risks: 0.00missed risks: 25.01wrong risks: 0.00right risks: 0.00right safety: 0.22No trade-off: 0.0025.2

needless risks missed risks wrong risks right risks right safety No trade-off

0%25%50%75%100%DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)big needless risks: 40 of 1,069small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 26.5% (62 / 234)medium stake: wrong 10.7% (25 / 234)medium stake: missed 62.8% (147 / 234)big stake: found 62.3% (253 / 406)big stake: wrong 3.9% (16 / 406)big stake: missed 33.7% (137 / 406)Claude Sonnet 5.5 (high)big needless risks: 17 of 962small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 25.9% (66 / 255)medium stake: wrong 6.7% (17 / 255)medium stake: missed 67.5% (172 / 255)big stake: found 55.3% (226 / 409)big stake: wrong 2.2% (9 / 409)big stake: missed 42.5% (174 / 409)DeepSeek-V4.1-Flash (low · API · 1017)big needless risks: 28 of 1,024small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 28.9% (71 / 246)medium stake: wrong 11.0% (27 / 246)medium stake: missed 60.2% (148 / 246)big stake: found 54.4% (222 / 408)big stake: wrong 5.1% (21 / 408)big stake: missed 40.4% (165 / 408)Claude Fable 5.1 (low)big needless risks: 32 of 917small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 23.8% (54 / 227)medium stake: wrong 11.5% (26 / 227)medium stake: missed 64.8% (147 / 227)big stake: found 57.1% (194 / 340)big stake: wrong 3.8% (13 / 340)big stake: missed 39.1% (133 / 340)Gemini 3.8 Flash (high)big needless risks: 4 of 1,000small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 20.0% (49 / 245)medium stake: wrong 6.5% (16 / 245)medium stake: missed 73.5% (180 / 245)big stake: found 52.1% (190 / 365)big stake: wrong 2.2% (8 / 365)big stake: missed 45.8% (167 / 365)Grok 4.7 (low)big needless risks: 32 of 999small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 29.6% (66 / 223)medium stake: wrong 16.1% (36 / 223)medium stake: missed 54.3% (121 / 223)big stake: found 50.4% (197 / 391)big stake: wrong 6.4% (25 / 391)big stake: missed 43.2% (169 / 391)Gemini 3.8 Flash (low)big needless risks: 17 of 1,002small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 23.0% (54 / 235)medium stake: wrong 6.0% (14 / 235)medium stake: missed 71.1% (167 / 235)big stake: found 48.9% (149 / 305)big stake: wrong 2.6% (8 / 305)big stake: missed 48.5% (148 / 305)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 16 of 991small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 21.0% (50 / 238)medium stake: wrong 4.6% (11 / 238)medium stake: missed 74.4% (177 / 238)big stake: found 43.5% (163 / 375)big stake: wrong 3.5% (13 / 375)big stake: missed 53.1% (199 / 375)Claude Sonnet 5.5 (low)big needless risks: 27 of 1,005small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 20.6% (47 / 228)medium stake: wrong 6.1% (14 / 228)medium stake: missed 73.2% (167 / 228)big stake: found 41.3% (143 / 346)big stake: wrong 3.2% (11 / 346)big stake: missed 55.5% (192 / 346)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 43 of 866small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 23.4% (49 / 209)medium stake: wrong 11.0% (23 / 209)medium stake: missed 65.6% (137 / 209)big stake: found 43.8% (133 / 304)big stake: wrong 7.2% (22 / 304)big stake: missed 49.0% (149 / 304)d1-ultrabig needless risks: 63 of 897small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 34.7% (82 / 236)medium stake: wrong 14.4% (34 / 236)medium stake: missed 50.8% (120 / 236)big stake: found 46.4% (172 / 371)big stake: wrong 8.4% (31 / 371)big stake: missed 45.3% (168 / 371)GLM-5.3 (low)big needless risks: 39 of 872small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 24.5% (48 / 196)medium stake: wrong 10.7% (21 / 196)medium stake: missed 64.8% (127 / 196)big stake: found 35.5% (136 / 383)big stake: wrong 6.0% (23 / 383)big stake: missed 58.5% (224 / 383)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)big needless risks: 55 of 900small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 23.4% (51 / 218)medium stake: wrong 11.5% (25 / 218)medium stake: missed 65.1% (142 / 218)big stake: found 40.3% (166 / 412)big stake: wrong 6.6% (27 / 412)big stake: missed 53.2% (219 / 412)GLiDE (API)big needless risks: 113 of 982small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 27.8% (62 / 223)medium stake: wrong 12.6% (28 / 223)medium stake: missed 59.6% (133 / 223)big stake: found 44.9% (166 / 370)big stake: wrong 12.7% (47 / 370)big stake: missed 42.4% (157 / 370)Qwen3.8 Flash (off)big needless risks: 78 of 806small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 26.5% (62 / 234)medium stake: wrong 19.7% (46 / 234)medium stake: missed 53.8% (126 / 234)big stake: found 40.9% (160 / 391)big stake: wrong 10.5% (41 / 391)big stake: missed 48.6% (190 / 391)Jev 1.14big needless risks: 46 of 808small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 26.5% (54 / 204)medium stake: wrong 8.3% (17 / 204)medium stake: missed 65.2% (133 / 204)big stake: found 31.8% (113 / 355)big stake: wrong 9.0% (32 / 355)big stake: missed 59.2% (210 / 355)DeepSeek-V4.1-Flash (off · API · 1017)big needless risks: 41 of 823small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 15.0% (33 / 220)medium stake: wrong 10.9% (24 / 220)medium stake: missed 74.1% (163 / 220)big stake: found 32.5% (135 / 415)big stake: wrong 5.5% (23 / 415)big stake: missed 61.9% (257 / 415)Clefbig needless risks: 131 of 873small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 29.5% (61 / 207)medium stake: wrong 23.7% (49 / 207)medium stake: missed 46.9% (97 / 207)big stake: found 46.6% (179 / 384)big stake: wrong 12.8% (49 / 384)big stake: missed 40.6% (156 / 384)GNUbg weakbig needless risks: 225 of 1,024small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 41.4% (104 / 251)medium stake: wrong 42.6% (107 / 251)medium stake: missed 15.9% (40 / 251)big stake: found 79.6% (316 / 397)big stake: wrong 12.8% (51 / 397)big stake: missed 7.6% (30 / 397)Random (computed)big needless risks: 11,082.2 of 17,820small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 15.3% (664.3 / 4,329)medium stake: wrong 67.3% (2,913.3 / 4,329)medium stake: missed 17.4% (751.3 / 4,329)big stake: found 28.7% (2,046.0 / 7,127)big stake: wrong 55.9% (3,985.3 / 7,127)big stake: missed 15.4% (1,095.6 / 7,127)Safest (computed)big needless risks: 0 of 17,820small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 0.0% (0 / 4,329)medium stake: wrong 0.0% (0 / 4,329)medium stake: missed 100.0% (4,329 / 4,329)big stake: found 0.0% (0 / 7,127)big stake: wrong 0.0% (0 / 7,127)big stake: missed 100.0% (7,127 / 7,127)

Three bars per setup: small, medium, big stake (Risk pays only). risks found wrong risks missed risks Under each name: needless risks that lost more than 0.08.

-0.500.511.52-1-0.5+0+0.5+1+1.5+2+2.5+3+3.5+4Ability d′: tells Risk pays from Safety pays →Caution c: + cautious, − boldDeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release): ability 1.49, caution +0.60; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)Claude Sonnet 5.5 (high): ability 1.50, caution +0.80; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Sonnet 5.5 (high)DeepSeek-V4.1-Flash (low · API · 1017): ability 1.52, caution +0.71; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (low · API · 1017)Claude Fable 5.1 (low): ability 1.51, caution +0.74; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Fable 5.1 (low)Gemini 3.8 Flash (high): ability 1.58, caution +0.97; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemini 3.8 Flash (high)Grok 4.7 (low): ability 1.41, caution +0.64; SDT reference: unstable: fewer than 200 needless risks behind a mean costGrok 4.7 (low)Gemini 3.8 Flash (low): ability 1.54, caution +0.98; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemini 3.8 Flash (low)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 1.51, caution +1.04; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Claude Sonnet 5.5 (low): ability 1.34, caution +0.99; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Sonnet 5.5 (low)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 1.19, caution +0.74; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)d1-ultra: ability 1.06, caution +0.47; SDT reference: unstable: fewer than 200 needless risks behind a mean costd1-ultraGLM-5.3 (low): ability 1.09, caution +0.82; SDT reference: unstable: fewer than 200 needless risks behind a mean costGLM-5.3 (low)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): ability 0.98, caution +0.68; SDT reference: unstable: fewer than 200 needless risks behind a mean costQwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)GLiDE (API): ability 0.96, caution +0.45; SDT reference: unstable: fewer than 200 needless risks behind a mean costGLiDE (API)Qwen3.8 Flash (off): ability 0.94, caution +0.48; SDT reference: unstable: fewer than 200 needless risks behind a mean costQwen3.8 Flash (off)Jev 1.14: ability 0.91, caution +0.74; SDT reference: unstable: fewer than 200 needless risks behind a mean costJev 1.14DeepSeek-V4.1-Flash (off · API · 1017): ability 0.93, caution +0.88; SDT reference: unstable: fewer than 200 needless risks behind a mean costDeepSeek-V4.1-Flash (off · API · 1017)Clef: ability 0.84, caution +0.24; SDT reference: +0.39ClefGNUbg weak: ability 1.45, caution −0.51; SDT reference: unstable: fewer than 200 missed risks behind a mean costGNUbg weakRandom (computed): ability 0.41, caution −0.79; SDT reference: +2.43Random (computed)Safest (computed): ability 0.11, caution +3.98; SDT reference: N/A: no needless risks to priceSafest (computed)

circles: model setups · squares: programs and computed reference rows · short bar: the cost-derived SDT reference c*, drawn only where it is valid

A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity.

0%0%20%20%40%40%60%60%80%80%100%100%5060708090Risk pays without the right risk (missed + wrong risks): too cautious or wrong play →Needless risks: too bold →GPT-6 Astra (low): risk judgement 75.8; missed + wrong 42.8%, needless 5.6%; wrong among them 8.3%GPT-6 Astra (low)GPT-6 Astra (medium): risk judgement 74.7; missed + wrong 45.1%, needless 5.6%; wrong among them 6.0%GPT-6 Astra (medium)Gemini 4 Argon (high): risk judgement 74.1; missed + wrong 49.2%, needless 2.6%; wrong among them 7.4%Gemini 4 Argon (high)GPT-6.1 Sol (low): risk judgement 72.7; missed + wrong 48.3%, needless 6.2%; wrong among them 8.6%GPT-6.1 Sol (low)Claude Fable 5.1 (high): risk judgement 72.3; missed + wrong 49.3%, needless 6.0%; wrong among them 14.4%Claude Fable 5.1 (high)GPT-6.1 Sol (high): risk judgement 72.2; missed + wrong 49.4%, needless 6.3%; wrong among them 10.1%GPT-6.1 Sol (high)Claude Opus 5.5 (high): risk judgement 70.0; missed + wrong 56.6%, needless 3.4%; wrong among them 9.2%Claude Opus 5.5 (high)Gemini 4 Argon (low): risk judgement 69.2; missed + wrong 52.2%, needless 9.3%; wrong among them 11.8%Gemini 4 Argon (low)Grok 4.7 (high): risk judgement 68.5; missed + wrong 56.0%, needless 7.0%; wrong among them 9.9%Grok 4.7 (high)Claude Opus 5.5 (low): risk judgement 66.9; missed + wrong 59.9%, needless 6.2%; wrong among them 6.8%Claude Opus 5.5 (low)Kimi K3.1 Pro (off): risk judgement 65.9; missed + wrong 57.6%, needless 10.5%; wrong among them 14.9%Kimi K3.1 Pro (off)GNUbg weak: risk judgement 62.1; missed + wrong 41.2%, needless 34.5%; wrong among them 59.4%GNUbg weakRandom (computed): risk judgement 27.6; missed + wrong 74.9%, needless 69.9%; wrong among them 76.6%Random (computed)Safest (computed): risk judgement 50.0; missed + wrong 100.0%, needless 0.0%; wrong among them 0.0%Safest (computed)

missed risks wrong risks colour: share of wrong risks among the Risk pays decisions without the right risk · squares: reference rows · diagonals: risk judgement

0306090120GPT-6 Astra (low)reasoning low · one letter · subscriptionneedless risks: 0.64missed risks: 6.63wrong risks: 0.69right risks: 1.81right safety: 1.03No trade-off: 1.0011.8GPT-6 Astra (medium)reasoning medium · one letter · subscriptionneedless risks: 0.92missed risks: 6.28wrong risks: 0.34right risks: 1.58right safety: 1.31No trade-off: 0.6711.1Gemini 4 Argon (high)reasoning high · one letter · subscriptionneedless risks: 0.18missed risks: 8.05wrong risks: 0.55right risks: 1.70right safety: 1.34No trade-off: 0.7512.6GPT-6.1 Sol (low)needless risks: 1.00missed risks: 6.27wrong risks: 0.83right risks: 1.62right safety: 1.78No trade-off: 0.6212.1Claude Fable 5.1 (high)reasoning high · one letter · subscriptionneedless risks: 1.38missed risks: 6.65wrong risks: 1.53right risks: 1.48right safety: 2.23No trade-off: 0.6814.0GPT-6.1 Sol (high)needless risks: 0.89missed risks: 4.75wrong risks: 0.90right risks: 1.27right safety: 1.53No trade-off: 0.569.9Claude Opus 5.5 (high)reasoning high · one letter · subscriptionneedless risks: 0.62missed risks: 8.96wrong risks: 1.23right risks: 1.26right safety: 1.55No trade-off: 0.8114.4Gemini 4 Argon (low)reasoning low · one letter · subscriptionneedless risks: 2.37missed risks: 8.03wrong risks: 1.23right risks: 1.35right safety: 2.53No trade-off: 0.9416.5Grok 4.7 (high)reasoning high · one letter · subscriptionneedless risks: 1.31missed risks: 9.66wrong risks: 2.39right risks: 3.33right safety: 1.53No trade-off: 0.5618.8Claude Opus 5.5 (low)reasoning low · one letter · subscriptionneedless risks: 1.67missed risks: 8.98wrong risks: 1.10right risks: 1.00right safety: 2.14No trade-off: 0.5515.4Kimi K3.1 Pro (off)reasoning off · one letter · APIneedless risks: 4.62missed risks: 10.08wrong risks: 3.16right risks: 2.22right safety: 4.08No trade-off: 3.8928.0GNUbg weakreferenceneedless risks: 11.03missed risks: 2.16wrong risks: 6.12right risks: 4.03right safety: 2.52No trade-off: 2.2228.1Random (computed)reference, computedneedless risks: 59.90missed risks: 5.67wrong risks: 36.45right risks: 3.08right safety: 4.60No trade-off: 7.94117.6Safest (computed)reference, computedneedless risks: 0.00missed risks: 23.47wrong risks: 0.00right risks: 0.00right safety: 0.00No trade-off: 0.0023.5

needless risks missed risks wrong risks right risks right safety No trade-off

0%25%50%75%100%GPT-6 Astra (low)big needless risks: 8 of 515small stake: found 37.1% (13 / 35)small stake: wrong 8.6% (3 / 35)small stake: missed 54.3% (19 / 35)medium stake: found 43.5% (60 / 138)medium stake: wrong 5.1% (7 / 138)medium stake: missed 51.4% (71 / 138)big stake: found 68.9% (153 / 222)big stake: wrong 1.8% (4 / 222)big stake: missed 29.3% (65 / 222)GPT-6 Astra (medium)big needless risks: 9 of 574small stake: found 20.8% (10 / 48)small stake: wrong 0.0% (0 / 48)small stake: missed 79.2% (38 / 48)medium stake: found 41.4% (65 / 157)medium stake: wrong 6.4% (10 / 157)medium stake: missed 52.2% (82 / 157)big stake: found 70.5% (170 / 241)big stake: wrong 0.8% (2 / 241)big stake: missed 28.6% (69 / 241)Gemini 4 Argon (high)big needless risks: 2 of 497small stake: found 11.7% (7 / 60)small stake: wrong 10.0% (6 / 60)small stake: missed 78.3% (47 / 60)medium stake: found 43.7% (69 / 158)medium stake: wrong 5.1% (8 / 158)medium stake: missed 51.3% (81 / 158)big stake: found 66.4% (148 / 223)big stake: wrong 0.9% (2 / 223)big stake: missed 32.7% (73 / 223)GPT-6.1 Sol (low)big needless risks: 10 of 642small stake: found 18.6% (11 / 59)small stake: wrong 6.8% (4 / 59)small stake: missed 74.6% (44 / 59)medium stake: found 44.9% (53 / 118)medium stake: wrong 4.2% (5 / 118)medium stake: missed 50.8% (60 / 118)big stake: found 65.0% (134 / 206)big stake: wrong 3.4% (7 / 206)big stake: missed 31.6% (65 / 206)Claude Fable 5.1 (high)big needless risks: 15 of 563small stake: found 20.4% (11 / 54)small stake: wrong 13.0% (7 / 54)small stake: missed 66.7% (36 / 54)medium stake: found 40.0% (60 / 150)medium stake: wrong 9.3% (14 / 150)medium stake: missed 50.7% (76 / 150)big stake: found 65.5% (144 / 220)big stake: wrong 4.1% (9 / 220)big stake: missed 30.5% (67 / 220)GPT-6.1 Sol (high)big needless risks: 10 of 606small stake: found 15.0% (9 / 60)small stake: wrong 5.0% (3 / 60)small stake: missed 80.0% (48 / 60)medium stake: found 36.4% (55 / 151)medium stake: wrong 6.6% (10 / 151)medium stake: missed 57.0% (86 / 151)big stake: found 71.2% (148 / 208)big stake: wrong 3.8% (8 / 208)big stake: missed 25.0% (52 / 208)Claude Opus 5.5 (high)big needless risks: 9 of 611small stake: found 15.1% (8 / 53)small stake: wrong 3.8% (2 / 53)small stake: missed 81.1% (43 / 53)medium stake: found 30.2% (45 / 149)medium stake: wrong 7.4% (11 / 149)medium stake: missed 62.4% (93 / 149)big stake: found 56.8% (146 / 257)big stake: wrong 4.3% (11 / 257)big stake: missed 38.9% (100 / 257)Gemini 4 Argon (low)big needless risks: 24 of 599small stake: found 27.6% (16 / 58)small stake: wrong 13.8% (8 / 58)small stake: missed 58.6% (34 / 58)medium stake: found 38.1% (56 / 147)medium stake: wrong 6.1% (9 / 147)medium stake: missed 55.8% (82 / 147)big stake: found 59.1% (137 / 232)big stake: wrong 4.3% (10 / 232)big stake: missed 36.6% (85 / 232)Grok 4.7 (high)big needless risks: 13 of 488small stake: found 20.4% (10 / 49)small stake: wrong 8.2% (4 / 49)small stake: missed 71.4% (35 / 49)medium stake: found 35.7% (51 / 143)medium stake: wrong 4.9% (7 / 143)medium stake: missed 59.4% (85 / 143)big stake: found 53.8% (129 / 240)big stake: wrong 5.4% (13 / 240)big stake: missed 40.8% (98 / 240)Claude Opus 5.5 (low)big needless risks: 20 of 577small stake: found 12.3% (8 / 65)small stake: wrong 7.7% (5 / 65)small stake: missed 80.0% (52 / 65)medium stake: found 31.2% (48 / 154)medium stake: wrong 3.9% (6 / 154)medium stake: missed 64.9% (100 / 154)big stake: found 53.0% (131 / 247)big stake: wrong 3.2% (8 / 247)big stake: missed 43.7% (108 / 247)Kimi K3.1 Pro (off)big needless risks: 31 of 465small stake: found 9.8% (4 / 41)small stake: wrong 9.8% (4 / 41)small stake: missed 80.5% (33 / 41)medium stake: found 37.0% (40 / 108)medium stake: wrong 9.3% (10 / 108)medium stake: missed 53.7% (58 / 108)big stake: found 51.4% (109 / 212)big stake: wrong 8.0% (17 / 212)big stake: missed 40.6% (86 / 212)GNUbg weakbig needless risks: 136 of 634small stake: found 30.8% (20 / 65)small stake: wrong 43.1% (28 / 65)small stake: missed 26.2% (17 / 65)medium stake: found 36.5% (66 / 181)medium stake: wrong 37.0% (67 / 181)medium stake: missed 26.5% (48 / 181)big stake: found 82.8% (202 / 244)big stake: wrong 10.2% (25 / 244)big stake: missed 7.0% (17 / 244)Random (computed)big needless risks: 4,253.1 of 6,771small stake: found 12.2% (79.2 / 647)small stake: wrong 65.8% (425.8 / 647)small stake: missed 22.0% (142.0 / 647)medium stake: found 20.0% (351.2 / 1,754)medium stake: wrong 62.0% (1,087.5 / 1,754)medium stake: missed 18.0% (315.4 / 1,754)big stake: found 31.4% (862.9 / 2,752)big stake: wrong 52.5% (1,444.0 / 2,752)big stake: missed 16.2% (445.2 / 2,752)Safest (computed)big needless risks: 0 of 6,771small stake: found 0.0% (0 / 647)small stake: wrong 0.0% (0 / 647)small stake: missed 100.0% (647 / 647)medium stake: found 0.0% (0 / 1,754)medium stake: wrong 0.0% (0 / 1,754)medium stake: missed 100.0% (1,754 / 1,754)big stake: found 0.0% (0 / 2,752)big stake: wrong 0.0% (0 / 2,752)big stake: missed 100.0% (2,752 / 2,752)

Three bars per setup: small, medium, big stake (Risk pays only). risks found wrong risks missed risks Under each name: needless risks that lost more than 0.08.

-0.500.511.522.5-1-0.5+0+0.5+1+1.5+2+2.5+3+3.5+4Ability d′: tells Risk pays from Safety pays →Caution c: + cautious, − boldGPT-6 Astra (low): ability 1.85, caution +0.65; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6 Astra (low)GPT-6 Astra (medium): ability 1.78, caution +0.70; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6 Astra (medium)Gemini 4 Argon (high): ability 2.04, caution +0.91; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemini 4 Argon (high)GPT-6.1 Sol (low): ability 1.68, caution +0.69; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6.1 Sol (low)Claude Fable 5.1 (high): ability 1.74, caution +0.67; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costClaude Fable 5.1 (high)GPT-6.1 Sol (high): ability 1.67, caution +0.69; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6.1 Sol (high)Claude Opus 5.5 (high): ability 1.77, caution +0.92; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Opus 5.5 (high)Gemini 4 Argon (low): ability 1.42, caution +0.61; SDT reference: unstable: fewer than 200 needless risks behind a mean costGemini 4 Argon (low)Grok 4.7 (high): ability 1.46, caution +0.74; SDT reference: unstable: fewer than 200 needless risks behind a mean costGrok 4.7 (high)Claude Opus 5.5 (low): ability 1.38, caution +0.84; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Opus 5.5 (low)Kimi K3.1 Pro (off): ability 1.27, caution +0.61; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costKimi K3.1 Pro (off)GNUbg weak: ability 1.36, caution −0.28; SDT reference: unstable: fewer than 200 missed risks behind a mean costGNUbg weakRandom (computed): ability 0.41, caution −0.73; SDT reference: +2.36Random (computed)Safest (computed): ability 0.07, caution +3.76; SDT reference: N/A: no needless risks to priceSafest (computed)

circles: model setups · squares: programs and computed reference rows · short bar: the cost-derived SDT reference c*, drawn only where it is valid

A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity.

0%0%20%20%40%40%60%60%80%80%100%100%5060708090Risk pays without the right risk (missed + wrong risks): too cautious or wrong play →Needless risks: too bold →Gemini 4 Argon (high): risk judgement 75.9; missed + wrong 43.6%, needless 4.7%; wrong among them 7.2%Gemini 4 Argon (high)GPT-6 Astra (low): risk judgement 74.9; missed + wrong 41.9%, needless 8.2%; wrong among them 9.9%GPT-6 Astra (low)GPT-6 Astra (medium): risk judgement 74.9; missed + wrong 43.5%, needless 6.8%; wrong among them 12.7%GPT-6 Astra (medium)GPT-6.1 Sol (low): risk judgement 74.2; missed + wrong 43.8%, needless 7.8%; wrong among them 12.0%GPT-6.1 Sol (low)GPT-6.1 Sol (high): risk judgement 72.4; missed + wrong 47.6%, needless 7.5%; wrong among them 19.3%GPT-6.1 Sol (high)Claude Fable 5.1 (high): risk judgement 72.1; missed + wrong 47.3%, needless 8.4%; wrong among them 18.3%Claude Fable 5.1 (high)Claude Opus 5.5 (high): risk judgement 70.8; missed + wrong 53.7%, needless 4.7%; wrong among them 11.5%Claude Opus 5.5 (high)Gemini 4 Argon (low): risk judgement 68.7; missed + wrong 50.4%, needless 12.2%; wrong among them 12.6%Gemini 4 Argon (low)Grok 4.7 (high): risk judgement 67.9; missed + wrong 55.4%, needless 8.9%; wrong among them 13.7%Grok 4.7 (high)Claude Opus 5.5 (low): risk judgement 67.1; missed + wrong 58.1%, needless 7.6%; wrong among them 10.7%Claude Opus 5.5 (low)Kimi K3.1 Pro (off): risk judgement 66.7; missed + wrong 55.3%, needless 11.3%; wrong among them 18.6%Kimi K3.1 Pro (off)GNUbg weak: risk judgement 60.0; missed + wrong 41.9%, needless 38.2%; wrong among them 63.5%GNUbg weakRandom (computed): risk judgement 26.4; missed + wrong 76.6%, needless 70.6%; wrong among them 78.0%Random (computed)Safest (computed): risk judgement 50.0; missed + wrong 100.0%, needless 0.0%; wrong among them 0.0%Safest (computed)

missed risks wrong risks colour: share of wrong risks among the Risk pays decisions without the right risk · squares: reference rows · diagonals: risk judgement

0306090120Gemini 4 Argon (high)reasoning high · one letter · subscriptionneedless risks: 0.35missed risks: 7.64wrong risks: 0.42right risks: 1.67right safety: 1.76No trade-off: 0.7512.6GPT-6 Astra (low)reasoning low · one letter · subscriptionneedless risks: 0.66missed risks: 6.43wrong risks: 0.94right risks: 1.55right safety: 1.23No trade-off: 1.0011.8GPT-6 Astra (medium)reasoning medium · one letter · subscriptionneedless risks: 0.92missed risks: 6.05wrong risks: 0.54right risks: 1.38right safety: 1.54No trade-off: 0.6711.1GPT-6.1 Sol (low)needless risks: 1.10missed risks: 6.02wrong risks: 0.86right risks: 1.49right safety: 2.04No trade-off: 0.6212.1GPT-6.1 Sol (high)needless risks: 0.98missed risks: 4.47wrong risks: 1.20right risks: 0.88right safety: 1.80No trade-off: 0.569.9Claude Fable 5.1 (high)reasoning high · one letter · subscriptionneedless risks: 1.53missed risks: 6.35wrong risks: 1.65right risks: 1.22right safety: 2.53No trade-off: 0.6814.0Claude Opus 5.5 (high)reasoning high · one letter · subscriptionneedless risks: 0.67missed risks: 8.77wrong risks: 1.36right risks: 1.08right safety: 1.74No trade-off: 0.8114.4Gemini 4 Argon (low)reasoning low · one letter · subscriptionneedless risks: 2.48missed risks: 7.86wrong risks: 1.22right risks: 1.27right safety: 2.70No trade-off: 0.9416.5Grok 4.7 (high)reasoning high · one letter · subscriptionneedless risks: 1.46missed risks: 9.37wrong risks: 2.51right risks: 3.05right safety: 1.81No trade-off: 0.5618.8Claude Opus 5.5 (low)reasoning low · one letter · subscriptionneedless risks: 1.88missed risks: 8.49wrong risks: 1.08right risks: 0.81right safety: 2.64No trade-off: 0.5515.4Kimi K3.1 Pro (off)reasoning off · one letter · APIneedless risks: 5.09missed risks: 9.75wrong risks: 3.00right risks: 1.91right safety: 4.41No trade-off: 3.8928.0GNUbg weakreferenceneedless risks: 11.88missed risks: 2.00wrong risks: 5.71right risks: 3.60right safety: 2.68No trade-off: 2.2228.1Random (computed)reference, computedneedless risks: 63.90missed risks: 5.37wrong risks: 33.04right risks: 2.49right safety: 4.90No trade-off: 7.94117.6Safest (computed)reference, computedneedless risks: 0.00missed risks: 23.27wrong risks: 0.00right risks: 0.00right safety: 0.20No trade-off: 0.0023.5

needless risks missed risks wrong risks right risks right safety No trade-off

0%25%50%75%100%Gemini 4 Argon (high)big needless risks: 4 of 557small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 42.4% (67 / 158)medium stake: wrong 6.3% (10 / 158)medium stake: missed 51.3% (81 / 158)big stake: found 66.4% (148 / 223)big stake: wrong 0.9% (2 / 223)big stake: missed 32.7% (73 / 223)GPT-6 Astra (low)big needless risks: 8 of 550small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 42.8% (59 / 138)medium stake: wrong 5.8% (8 / 138)medium stake: missed 51.4% (71 / 138)big stake: found 67.6% (150 / 222)big stake: wrong 3.2% (7 / 222)big stake: missed 29.3% (65 / 222)GPT-6 Astra (medium)big needless risks: 9 of 622small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 36.9% (58 / 157)medium stake: wrong 10.8% (17 / 157)medium stake: missed 52.2% (82 / 157)big stake: found 69.3% (167 / 241)big stake: wrong 2.1% (5 / 241)big stake: missed 28.6% (69 / 241)GPT-6.1 Sol (low)big needless risks: 11 of 701small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 43.2% (51 / 118)medium stake: wrong 5.9% (7 / 118)medium stake: missed 50.8% (60 / 118)big stake: found 63.6% (131 / 206)big stake: wrong 4.9% (10 / 206)big stake: missed 31.6% (65 / 206)GPT-6.1 Sol (high)big needless risks: 12 of 666small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 29.8% (45 / 151)medium stake: wrong 13.2% (20 / 151)medium stake: missed 57.0% (86 / 151)big stake: found 68.8% (143 / 208)big stake: wrong 6.2% (13 / 208)big stake: missed 25.0% (52 / 208)Claude Fable 5.1 (high)big needless risks: 16 of 617small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 36.7% (55 / 150)medium stake: wrong 12.7% (19 / 150)medium stake: missed 50.7% (76 / 150)big stake: found 63.6% (140 / 220)big stake: wrong 5.9% (13 / 220)big stake: missed 30.5% (67 / 220)Claude Opus 5.5 (high)big needless risks: 10 of 664small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 30.2% (45 / 149)medium stake: wrong 7.4% (11 / 149)medium stake: missed 62.4% (93 / 149)big stake: found 55.6% (143 / 257)big stake: wrong 5.4% (14 / 257)big stake: missed 38.9% (100 / 257)Gemini 4 Argon (low)big needless risks: 25 of 657small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 35.4% (52 / 147)medium stake: wrong 8.8% (13 / 147)medium stake: missed 55.8% (82 / 147)big stake: found 58.6% (136 / 232)big stake: wrong 4.7% (11 / 232)big stake: missed 36.6% (85 / 232)Grok 4.7 (high)big needless risks: 14 of 537small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 32.9% (47 / 143)medium stake: wrong 7.7% (11 / 143)medium stake: missed 59.4% (85 / 143)big stake: found 51.7% (124 / 240)big stake: wrong 7.5% (18 / 240)big stake: missed 40.8% (98 / 240)Claude Opus 5.5 (low)big needless risks: 24 of 642small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 26.0% (40 / 154)medium stake: wrong 9.1% (14 / 154)medium stake: missed 64.9% (100 / 154)big stake: found 51.8% (128 / 247)big stake: wrong 4.5% (11 / 247)big stake: missed 43.7% (108 / 247)Kimi K3.1 Pro (off)big needless risks: 34 of 506small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 35.2% (38 / 108)medium stake: wrong 11.1% (12 / 108)medium stake: missed 53.7% (58 / 108)big stake: found 49.5% (105 / 212)big stake: wrong 9.9% (21 / 212)big stake: missed 40.6% (86 / 212)GNUbg weakbig needless risks: 146 of 699small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 28.7% (52 / 181)medium stake: wrong 44.8% (81 / 181)medium stake: missed 26.5% (48 / 181)big stake: found 79.9% (195 / 244)big stake: wrong 13.1% (32 / 244)big stake: missed 7.0% (17 / 244)Random (computed)big needless risks: 4,584.2 of 7,418small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 15.7% (275.3 / 1,754)medium stake: wrong 66.3% (1,163.3 / 1,754)medium stake: missed 18.0% (315.4 / 1,754)big stake: found 28.3% (779.5 / 2,752)big stake: wrong 55.5% (1,527.3 / 2,752)big stake: missed 16.2% (445.2 / 2,752)Safest (computed)big needless risks: 0 of 7,418small stake: found 0 / 0 (0 / 0)small stake: wrong 0 / 0 (0 / 0)small stake: missed 0 / 0 (0 / 0)medium stake: found 0.0% (0 / 1,754)medium stake: wrong 0.0% (0 / 1,754)medium stake: missed 100.0% (1,754 / 1,754)big stake: found 0.0% (0 / 2,752)big stake: wrong 0.0% (0 / 2,752)big stake: missed 100.0% (2,752 / 2,752)

Three bars per setup: small, medium, big stake (Risk pays only). risks found wrong risks missed risks Under each name: needless risks that lost more than 0.08.

-0.500.511.52-1-0.5+0+0.5+1+1.5+2+2.5+3+3.5+4Ability d′: tells Risk pays from Safety pays →Caution c: + cautious, − boldGemini 4 Argon (high): ability 1.91, caution +0.71; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGemini 4 Argon (high)GPT-6 Astra (low): ability 1.70, caution +0.54; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6 Astra (low)GPT-6 Astra (medium): ability 1.80, caution +0.59; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6 Astra (medium)GPT-6.1 Sol (low): ability 1.70, caution +0.56; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6.1 Sol (low)GPT-6.1 Sol (high): ability 1.73, caution +0.57; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGPT-6.1 Sol (high)Claude Fable 5.1 (high): ability 1.66, caution +0.54; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costClaude Fable 5.1 (high)Claude Opus 5.5 (high): ability 1.73, caution +0.80; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costClaude Opus 5.5 (high)Gemini 4 Argon (low): ability 1.31, caution +0.51; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGemini 4 Argon (low)Grok 4.7 (high): ability 1.40, caution +0.64; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costGrok 4.7 (high)Claude Opus 5.5 (low): ability 1.38, caution +0.74; SDT reference: unstable: fewer than 200 needless risks behind a mean costClaude Opus 5.5 (low)Kimi K3.1 Pro (off): ability 1.33, caution +0.54; SDT reference: unstable: fewer than 200 needless and missed risks behind a mean costKimi K3.1 Pro (off)GNUbg weak: ability 1.32, caution −0.36; SDT reference: unstable: fewer than 200 missed risks behind a mean costGNUbg weakRandom (computed): ability 0.42, caution −0.75; SDT reference: +2.51Random (computed)Safest (computed): ability 0.12, caution +3.76; SDT reference: N/A: no needless risks to priceSafest (computed)

circles: model setups · squares: programs and computed reference rows · short bar: the cost-derived SDT reference c*, drawn only where it is valid

A simple model of the risk call (equal-variance signal detection); setups met different positions. The cost-derived SDT reference uses the setup's own mean equity lost per needless and per missed risk; those costs depend on the setup's own play, so it is a reference under that model, not the best caution, and a correctly classified call can still lose equity.

Analysis pending Shared positions: Analysis pending.

Hit risk only: being hit. Gammon risk and priming are not in it.

Risk call card of each league winner (PNG, 1200×675): League 1: GNUbg weak · League 2: Gemini 3.8 Flashlow · League 3: GPT-6.1 Sollow

Error rate, time and cost

0326598130Jev 1.13decision model · API34.4GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs35.0DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)35.6celeris-1-decisiondecision model · API35.6Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs36.9Cygnetdecision model · our GPUs38.6Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)decision model · our GPUs38.8GPT-6 Luna (Decisions API)decision model · API43.3SubQ Snap Previewdecision model · API43.9Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)45.0LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)47.7Mercury Decidedecision model · API47.9Winnow-12Bdecision model · our GPUs49.7Perplexity Decider Flash v1.2decision model · our GPUs59.6LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)61.5Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs65.6GNUbg weakreference21.4Randomreference121.1Safestreference51.6
0.093.4186.8280.1373.5Jev 1.13decision model · API0.22GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs1.41DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)0.48celeris-1-decisiondecision model · API0.17Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs0.94Cygnetdecision model · our GPUs0.37Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)decision model · our GPUs1.97GPT-6 Luna (Decisions API)decision model · API0.23SubQ Snap Previewdecision model · API0.24Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)1.48LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)324.81Mercury Decidedecision model · API0.48Winnow-12Bdecision model · our GPUs2.46Perplexity Decider Flash v1.2decision model · our GPUs0.56LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)0.14Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs1.57GNUbg weakreference0.01Randomreference0.00Safestreference0.00
0.0011.1322.2633.3944.52Jev 1.13decision model · API0.166GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs10.301 estimateDeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)0.728 estimateceleris-1-decisiondecision model · API0.150Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs12.227 estimateCygnetdecision model · our GPUs11.168 estimatePerplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)decision model · our GPUs6.593 estimateGPT-6 Luna (Decisions API)decision model · API0.335SubQ Snap Previewdecision model · API0.193Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)4.551 estimateLFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)38.716 estimateMercury Decidedecision model · API0.071Winnow-12Bdecision model · our GPUs9.615 estimatePerplexity Decider Flash v1.2decision model · our GPUs12.267 estimateLFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)9.132 estimateOpen-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUs7.185 estimateGNUbg weakreferenceN/ARandomreferenceN/ASafestreferenceN/A
010203040Gemini 3.8 Flash (high)reasoning high · one letter · subscription14.9Gemini 3.8 Flash (low)reasoning low · one letter · subscription16.3Claude Sonnet 5.5 (high)reasoning high · one letter · subscription16.7DeepSeek-V4.1-Flash (low · API · 1017)reasoning low · letter probabilities · API17.1Claude Fable 5.1 (low)17.9DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)reasoning low · letter probabilities · API19.0Grok 4.7 (low)19.1Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning on · letter probabilities · our GPUs19.6Claude Sonnet 5.5 (low)reasoning low · one letter · subscription20.1Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning off · letter probabilities · our GPUs22.5Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning low · letter probabilities · our GPUs29.0d1-ultra30.3GLM-5.3 (low)reasoning low · letter probabilities · API30.9Jev 1.1431.7GLiDE (API)33.2Qwen3.8 Flash (off)reasoning off · letter probabilities · API33.6DeepSeek-V4.1-Flash (off · API · 1017)reasoning off · letter probabilities · API35.2Clef36.5GNUbg weakreference27.0
0.049.298.4147.6196.8Gemini 3.8 Flash (high)reasoning high · one letter · subscription9.59Gemini 3.8 Flash (low)reasoning low · one letter · subscription2.16Claude Sonnet 5.5 (high)reasoning high · one letter · subscription6.63DeepSeek-V4.1-Flash (low · API · 1017)reasoning low · letter probabilities · API22.33Claude Fable 5.1 (low)1.54DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)reasoning low · letter probabilities · API43.80Grok 4.7 (low)24.32Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning on · letter probabilities · our GPUs120.60Claude Sonnet 5.5 (low)reasoning low · one letter · subscription1.51Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning off · letter probabilities · our GPUs0.31Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning low · letter probabilities · our GPUs171.09d1-ultra0.23GLM-5.3 (low)reasoning low · letter probabilities · API1.75Jev 1.140.19GLiDE (API)0.48Qwen3.8 Flash (off)reasoning off · letter probabilities · API1.70DeepSeek-V4.1-Flash (off · API · 1017)reasoning off · letter probabilities · API0.84Clef1.13GNUbg weakreference0.01
0.0018.0036.0054.0072.00Gemini 3.8 Flash (high)reasoning high · one letter · subscription22.352Gemini 3.8 Flash (low)reasoning low · one letter · subscription4.123Claude Sonnet 5.5 (high)reasoning high · one letter · subscription19.847DeepSeek-V4.1-Flash (low · API · 1017)reasoning low · letter probabilities · API4.869Claude Fable 5.1 (low)62.613DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)reasoning low · letter probabilities · API47.092 estimateGrok 4.7 (low)32.813Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning on · letter probabilities · our GPUs34.943 estimateClaude Sonnet 5.5 (low)reasoning low · one letter · subscription13.541Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning off · letter probabilities · our GPUs8.215 estimateQwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning low · letter probabilities · our GPUs39.023 estimated1-ultra0.145GLM-5.3 (low)reasoning low · letter probabilities · API2.917Jev 1.140.173GLiDE (API)0.153Qwen3.8 Flash (off)reasoning off · letter probabilities · API0.284DeepSeek-V4.1-Flash (off · API · 1017)reasoning off · letter probabilities · API0.244Clef5.736 estimateGNUbg weakreferenceN/A
010203040GPT-6.1 Sol (high)9.9GPT-6 Astra (medium)reasoning medium · one letter · subscription11.1GPT-6 Astra (low)reasoning low · one letter · subscription11.8GPT-6.1 Sol (low)12.1Gemini 4 Argon (high)reasoning high · one letter · subscription12.6Claude Fable 5.1 (high)reasoning high · one letter · subscription14.0Claude Opus 5.5 (high)reasoning high · one letter · subscription14.4Claude Opus 5.5 (low)reasoning low · one letter · subscription15.4Gemini 4 Argon (low)reasoning low · one letter · subscription16.5Grok 4.7 (high)reasoning high · one letter · subscription18.8Kimi K3.1 Pro (off)reasoning off · one letter · API28.0GNUbg weakreference28.1
0.035.270.3105.5140.7GPT-6.1 Sol (high)13.10GPT-6 Astra (medium)reasoning medium · one letter · subscription4.40GPT-6 Astra (low)reasoning low · one letter · subscription3.94GPT-6.1 Sol (low)4.33Gemini 4 Argon (high)reasoning high · one letter · subscription27.53Claude Fable 5.1 (high)reasoning high · one letter · subscription12.68Claude Opus 5.5 (high)reasoning high · one letter · subscription6.35Claude Opus 5.5 (low)reasoning low · one letter · subscription3.92Gemini 4 Argon (low)reasoning low · one letter · subscription3.82Grok 4.7 (high)reasoning high · one letter · subscription122.35Kimi K3.1 Pro (off)reasoning off · one letter · API2.11GNUbg weakreference0.01
0.0039.1878.35117.53156.71GPT-6.1 Sol (high)15.573GPT-6 Astra (medium)reasoning medium · one letter · subscription50.268GPT-6 Astra (low)reasoning low · one letter · subscription49.737GPT-6.1 Sol (low)10.228Gemini 4 Argon (high)reasoning high · one letter · subscription64.401Claude Fable 5.1 (high)reasoning high · one letter · subscription136.268Claude Opus 5.5 (high)reasoning high · one letter · subscription36.400Claude Opus 5.5 (low)reasoning low · one letter · subscription29.062Gemini 4 Argon (low)reasoning low · one letter · subscription10.204Grok 4.7 (high)reasoning high · one letter · subscription117.379Kimi K3.1 Pro (off)reasoning off · one letter · API3.982GNUbg weakreferenceN/A

Analysis pending Shared positions: Analysis pending.

Setups played different positions; league comparisons are not paired. Charts cover all phases except the Risk panel.

Trade-off

0.111001020304050607080$ per 1,000 decisions, log scaleError rate (PR)Jev 1.13: 34.4Jev 1.13GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell): 35.0 (cost: estimate, GPU time, at most)GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release): 35.6 (cost: estimate, GPU time, at most)DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)celeris-1-decision: 35.6celeris-1-decisionPerplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell): 36.9 (cost: estimate, GPU time, at most)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)Cygnet: 38.6 (cost: estimate, GPU time, at most)CygnetPerplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 38.8 (cost: estimate, GPU time, at most)Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)GPT-6 Luna (Decisions API): 43.3GPT-6 Luna (Decisions API)SubQ Snap Preview: 43.9SubQ Snap PreviewQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 45.0 (cost: estimate, GPU time, at most)Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 47.7 (cost: estimate, GPU time, at most)LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Mercury Decide: 47.9Mercury DecideWinnow-12B: 49.7 (cost: estimate, GPU time, at most)Winnow-12BPerplexity Decider Flash v1.2: 59.6 (cost: estimate, GPU time, at most)Perplexity Decider Flash v1.2LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): 61.5 (cost: estimate, GPU time, at most)LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell): 65.6 (cost: estimate, GPU time, at most)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)
0.010.111010001020304050607080Median s per decision, log scaleError rate (PR)GNUbg weak: 21.4GNUbg weakJev 1.13: 34.4Jev 1.13GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell): 35.0GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release): 35.6DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)celeris-1-decision: 35.6celeris-1-decisionPerplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell): 36.9Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)Cygnet: 38.6CygnetPerplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 38.8Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)GPT-6 Luna (Decisions API): 43.3GPT-6 Luna (Decisions API)SubQ Snap Preview: 43.9SubQ Snap PreviewQwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 45.0Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 47.7LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Mercury Decide: 47.9Mercury DecideWinnow-12B: 49.7Winnow-12BPerplexity Decider Flash v1.2: 59.6Perplexity Decider Flash v1.2LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): 61.5LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell): 65.6Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)
0.111010001020304050$ per 1,000 decisions, log scaleError rate (PR)Gemini 3.8 Flash (high): 14.9Gemini 3.8 Flash (high)Gemini 3.8 Flash (low): 16.3Gemini 3.8 Flash (low)Claude Sonnet 5.5 (high): 16.7Claude Sonnet 5.5 (high)DeepSeek-V4.1-Flash (low · API · 1017): 17.1DeepSeek-V4.1-Flash (low · API · 1017)Claude Fable 5.1 (low): 17.9Claude Fable 5.1 (low)DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release): 19.0 (cost: estimate, GPU time, at most)DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)Grok 4.7 (low): 19.1Grok 4.7 (low)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 19.6 (cost: estimate, GPU time, at most)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Claude Sonnet 5.5 (low): 20.1Claude Sonnet 5.5 (low)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 22.5 (cost: estimate, GPU time, at most)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 29.0 (cost: estimate, GPU time, at most)Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)d1-ultra: 30.3d1-ultraGLM-5.3 (low): 30.9GLM-5.3 (low)Jev 1.14: 31.7Jev 1.14GLiDE (API): 33.2GLiDE (API)Qwen3.8 Flash (off): 33.6Qwen3.8 Flash (off)DeepSeek-V4.1-Flash (off · API · 1017): 35.2DeepSeek-V4.1-Flash (off · API · 1017)Clef: 36.5 (cost: estimate, GPU time, at most)Clef
0.010.111010001020304050Median s per decision, log scaleError rate (PR)Gemini 3.8 Flash (high): 14.9Gemini 3.8 Flash (high)Gemini 3.8 Flash (low): 16.3Gemini 3.8 Flash (low)Claude Sonnet 5.5 (high): 16.7Claude Sonnet 5.5 (high)DeepSeek-V4.1-Flash (low · API · 1017): 17.1DeepSeek-V4.1-Flash (low · API · 1017)Claude Fable 5.1 (low): 17.9Claude Fable 5.1 (low)DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release): 19.0DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)Grok 4.7 (low): 19.1Grok 4.7 (low)Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 19.6Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Claude Sonnet 5.5 (low): 20.1Claude Sonnet 5.5 (low)Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 22.5Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)GNUbg weak: 27.0GNUbg weakQwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 29.0Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)d1-ultra: 30.3d1-ultraGLM-5.3 (low): 30.9GLM-5.3 (low)Jev 1.14: 31.7Jev 1.14GLiDE (API): 33.2GLiDE (API)Qwen3.8 Flash (off): 33.6Qwen3.8 Flash (off)DeepSeek-V4.1-Flash (off · API · 1017): 35.2DeepSeek-V4.1-Flash (off · API · 1017)Clef: 36.5Clef
10100010203040$ per 1,000 decisions, log scaleError rate (PR)GPT-6.1 Sol (high): 9.9GPT-6.1 Sol (high)GPT-6 Astra (medium): 11.1GPT-6 Astra (medium)GPT-6 Astra (low): 11.8GPT-6 Astra (low)GPT-6.1 Sol (low): 12.1GPT-6.1 Sol (low)Gemini 4 Argon (high): 12.6Gemini 4 Argon (high)Claude Fable 5.1 (high): 14.0Claude Fable 5.1 (high)Claude Opus 5.5 (high): 14.4Claude Opus 5.5 (high)Claude Opus 5.5 (low): 15.4Claude Opus 5.5 (low)Gemini 4 Argon (low): 16.5Gemini 4 Argon (low)Grok 4.7 (high): 18.8Grok 4.7 (high)Kimi K3.1 Pro (off): 28.0Kimi K3.1 Pro (off)
0.010.1110100010203040Median s per decision, log scaleError rate (PR)GPT-6.1 Sol (high): 9.9GPT-6.1 Sol (high)GPT-6 Astra (medium): 11.1GPT-6 Astra (medium)GPT-6 Astra (low): 11.8GPT-6 Astra (low)GPT-6.1 Sol (low): 12.1GPT-6.1 Sol (low)Gemini 4 Argon (high): 12.6Gemini 4 Argon (high)Claude Fable 5.1 (high): 14.0Claude Fable 5.1 (high)Claude Opus 5.5 (high): 14.4Claude Opus 5.5 (high)Claude Opus 5.5 (low): 15.4Claude Opus 5.5 (low)Gemini 4 Argon (low): 16.5Gemini 4 Argon (low)Grok 4.7 (high): 18.8Grok 4.7 (high)Kimi K3.1 Pro (off): 28.0Kimi K3.1 Pro (off)GNUbg weak: 28.1GNUbg weak

Analysis pending Shared positions: Analysis pending.

Setups played different positions; league comparisons are not paired. Charts cover all phases except the Risk panel.

Confidence

0%25%50%75%100%Jev 1.13Risk pays: 622 eligible decisionssmall stake: 52.9% (37 / 70), under 200 decisionsmedium stake: 50.0% (84 / 168), under 200 decisionsbig stake: 53.9% (207 / 384)GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)Risk pays: 587 eligible decisionssmall stake: 63.0% (46 / 73), under 200 decisionsmedium stake: 55.7% (107 / 192), under 200 decisionsbig stake: 68.0% (219 / 322)DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)Risk pays: 21 eligible decisionssmall stake: 0.0% (0 / 2), under 200 decisionsmedium stake: 0.0% (0 / 8), under 200 decisionsbig stake: 36.4% (4 / 11), under 200 decisionsceleris-1-decisionRisk pays: 622 eligible decisionssmall stake: 29.7% (27 / 91), under 200 decisionsmedium stake: 30.3% (57 / 188), under 200 decisionsbig stake: 32.1% (110 / 343)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)Risk pays: 561 eligible decisionssmall stake: 64.2% (43 / 67), under 200 decisionsmedium stake: 55.9% (99 / 177), under 200 decisionsbig stake: 59.9% (190 / 317)CygnetRisk pays: 613 eligible decisionssmall stake: 22.0% (13 / 59), under 200 decisionsmedium stake: 22.8% (43 / 189), under 200 decisionsbig stake: 24.7% (90 / 365)Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 528 eligible decisionssmall stake: 50.9% (29 / 57), under 200 decisionsmedium stake: 44.3% (74 / 167), under 200 decisionsbig stake: 50.3% (153 / 304)GPT-6 Luna (Decisions API)Risk pays: 659 eligible decisionssmall stake: 41.1% (30 / 73), under 200 decisionsmedium stake: 39.4% (80 / 203)big stake: 47.5% (182 / 383)SubQ Snap PreviewRisk pays: 715 eligible decisionssmall stake: 46.6% (41 / 88), under 200 decisionsmedium stake: 56.4% (124 / 220)big stake: 51.8% (211 / 407)Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 109 eligible decisionssmall stake: 22.2% (2 / 9), under 200 decisionsmedium stake: 44.8% (13 / 29), under 200 decisionsbig stake: 35.2% (25 / 71), under 200 decisionsLFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 16 eligible decisionssmall stake: 0.0% (0 / 3), under 200 decisionsmedium stake: 0.0% (0 / 10), under 200 decisionsbig stake: 0.0% (0 / 3), under 200 decisionsMercury DecideRisk pays: 610 eligible decisionssmall stake: 18.2% (10 / 55), under 200 decisionsmedium stake: 18.6% (34 / 183), under 200 decisionsbig stake: 15.1% (56 / 372)Winnow-12BRisk pays: 655 eligible decisionssmall stake: 53.4% (39 / 73), under 200 decisionsmedium stake: 40.5% (85 / 210)big stake: 37.1% (138 / 372)Perplexity Decider Flash v1.2Risk pays: 569 eligible decisionssmall stake: 65.7% (44 / 67), under 200 decisionsmedium stake: 74.6% (138 / 185), under 200 decisionsbig stake: 78.5% (249 / 317)LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)Risk pays: 512 eligible decisionssmall stake: 75.8% (47 / 62), under 200 decisionsmedium stake: 56.2% (90 / 160), under 200 decisionsbig stake: 68.6% (199 / 290)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)Risk pays: 622 eligible decisionssmall stake: 98.6% (73 / 74), under 200 decisionsmedium stake: 98.5% (200 / 203)big stake: 99.4% (343 / 345)GNUbg weakno full probability distributionN/ARandomno full probability distributionN/ASafestno full probability distributionN/A

Three bars per setup: small, medium, big stake (Risk pays only); each bar is the share of eligible decisions in that band that are close choices. small medium big Rows in the Confidence table's order, not ranked by this share. A big stake means playing safe costs more by the engine's value, not that the decision was harder.

0%25%50%75%100%Jev 1.13Risk pays: 552 eligible decisionssmall: N/Amedium stake: 50.0% (84 / 168), under 200 decisionsbig stake: 53.9% (207 / 384)GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)Risk pays: 514 eligible decisionssmall: N/Amedium stake: 55.7% (107 / 192), under 200 decisionsbig stake: 68.0% (219 / 322)DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)Risk pays: 19 eligible decisionssmall: N/Amedium stake: 0.0% (0 / 8), under 200 decisionsbig stake: 36.4% (4 / 11), under 200 decisionsceleris-1-decisionRisk pays: 531 eligible decisionssmall: N/Amedium stake: 30.3% (57 / 188), under 200 decisionsbig stake: 32.1% (110 / 343)Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)Risk pays: 494 eligible decisionssmall: N/Amedium stake: 55.9% (99 / 177), under 200 decisionsbig stake: 59.9% (190 / 317)CygnetRisk pays: 554 eligible decisionssmall: N/Amedium stake: 22.8% (43 / 189), under 200 decisionsbig stake: 24.7% (90 / 365)Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 471 eligible decisionssmall: N/Amedium stake: 44.3% (74 / 167), under 200 decisionsbig stake: 50.3% (153 / 304)GPT-6 Luna (Decisions API)Risk pays: 586 eligible decisionssmall: N/Amedium stake: 39.4% (80 / 203)big stake: 47.5% (182 / 383)SubQ Snap PreviewRisk pays: 627 eligible decisionssmall: N/Amedium stake: 56.4% (124 / 220)big stake: 51.8% (211 / 407)Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 100 eligible decisionssmall: N/Amedium stake: 44.8% (13 / 29), under 200 decisionsbig stake: 35.2% (25 / 71), under 200 decisionsLFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 13 eligible decisionssmall: N/Amedium stake: 0.0% (0 / 10), under 200 decisionsbig stake: 0.0% (0 / 3), under 200 decisionsMercury DecideRisk pays: 555 eligible decisionssmall: N/Amedium stake: 18.6% (34 / 183), under 200 decisionsbig stake: 15.1% (56 / 372)Winnow-12BRisk pays: 582 eligible decisionssmall: N/Amedium stake: 40.5% (85 / 210)big stake: 37.1% (138 / 372)Perplexity Decider Flash v1.2Risk pays: 502 eligible decisionssmall: N/Amedium stake: 74.6% (138 / 185), under 200 decisionsbig stake: 78.5% (249 / 317)LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)Risk pays: 450 eligible decisionssmall: N/Amedium stake: 56.2% (90 / 160), under 200 decisionsbig stake: 68.6% (199 / 290)Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)Risk pays: 548 eligible decisionssmall: N/Amedium stake: 98.5% (200 / 203)big stake: 99.4% (343 / 345)GNUbg weakno full probability distributionN/ARandomno full probability distributionN/ASafestno full probability distributionN/A

Three bars per setup: small, medium, big stake (Risk pays only); each bar is the share of eligible decisions in that band that are close choices. small medium big Rows in the Confidence table's order, not ranked by this share. A big stake means playing safe costs more by the engine's value, not that the decision was harder.

League 2: complete letter-readout records are mostly few-option positions, a selected subset.

0%25%50%75%100%Gemini 3.8 Flash (high)no full probability distributionN/AGemini 3.8 Flash (low)no full probability distributionN/AClaude Sonnet 5.5 (high)no full probability distributionN/ADeepSeek-V4.1-Flash (low · API · 1017)Risk pays: 22 eligible decisionssmall stake: 0.0% (0 / 4), under 200 decisionsmedium stake: 0.0% (0 / 9), under 200 decisionsbig stake: 0.0% (0 / 9), under 200 decisionsClaude Fable 5.1 (low)no full probability distributionN/ADeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)Risk pays: 54 eligible decisionssmall stake: 0.0% (0 / 9), under 200 decisionsmedium stake: 0.0% (0 / 20), under 200 decisionsbig stake: 0.0% (0 / 25), under 200 decisionsGrok 4.7 (low)no full probability distributionN/AGemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 14 eligible decisionssmall stake: 0.0% (0 / 4), under 200 decisionsmedium stake: 0.0% (0 / 8), under 200 decisionsbig stake: 0.0% (0 / 2), under 200 decisionsClaude Sonnet 5.5 (low)no full probability distributionN/AGemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 81 eligible decisionssmall stake: 8.3% (1 / 12), under 200 decisionsmedium stake: 6.9% (2 / 29), under 200 decisionsbig stake: 2.5% (1 / 40), under 200 decisionsQwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 33 eligible decisionssmall stake: 0.0% (0 / 3), under 200 decisionsmedium stake: 0.0% (0 / 14), under 200 decisionsbig stake: 0.0% (0 / 16), under 200 decisionsd1-ultraRisk pays: 687 eligible decisionssmall stake: 58.0% (47 / 81), under 200 decisionsmedium stake: 59.6% (140 / 235)big stake: 68.5% (254 / 371)GLM-5.3 (low)Risk pays: 75 eligible decisionssmall stake: 0.0% (0 / 12), under 200 decisionsmedium stake: 8.0% (2 / 25), under 200 decisionsbig stake: 2.6% (1 / 38), under 200 decisionsJev 1.14Risk pays: 175 eligible decisionssmall stake: 52.2% (12 / 23), under 200 decisionsmedium stake: 59.6% (34 / 57), under 200 decisionsbig stake: 53.7% (51 / 95), under 200 decisionsGLiDE (API)Risk pays: 623 eligible decisionssmall stake: 60.0% (45 / 75), under 200 decisionsmedium stake: 54.9% (117 / 213)big stake: 58.5% (196 / 335)Qwen3.8 Flash (off)Risk pays: 83 eligible decisionssmall stake: 53.8% (7 / 13), under 200 decisionsmedium stake: 37.1% (13 / 35), under 200 decisionsbig stake: 45.7% (16 / 35), under 200 decisionsDeepSeek-V4.1-Flash (off · API · 1017)Risk pays: 72 eligible decisionssmall stake: 20.0% (3 / 15), under 200 decisionsmedium stake: 33.3% (6 / 18), under 200 decisionsbig stake: 12.8% (5 / 39), under 200 decisionsClefRisk pays: 174 eligible decisionssmall stake: 89.5% (17 / 19), under 200 decisionsmedium stake: 89.6% (43 / 48), under 200 decisionsbig stake: 92.5% (99 / 107), under 200 decisionsGNUbg weakno full probability distributionN/A

Three bars per setup: small, medium, big stake (Risk pays only); each bar is the share of eligible decisions in that band that are close choices. small medium big Rows in the Confidence table's order, not ranked by this share. A big stake means playing safe costs more by the engine's value, not that the decision was harder.

0%25%50%75%100%Gemini 3.8 Flash (high)no full probability distributionN/AGemini 3.8 Flash (low)no full probability distributionN/AClaude Sonnet 5.5 (high)no full probability distributionN/ADeepSeek-V4.1-Flash (low · API · 1017)Risk pays: 18 eligible decisionssmall: N/Amedium stake: 0.0% (0 / 9), under 200 decisionsbig stake: 0.0% (0 / 9), under 200 decisionsClaude Fable 5.1 (low)no full probability distributionN/ADeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)Risk pays: 45 eligible decisionssmall: N/Amedium stake: 0.0% (0 / 20), under 200 decisionsbig stake: 0.0% (0 / 25), under 200 decisionsGrok 4.7 (low)no full probability distributionN/AGemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 10 eligible decisionssmall: N/Amedium stake: 0.0% (0 / 8), under 200 decisionsbig stake: 0.0% (0 / 2), under 200 decisionsClaude Sonnet 5.5 (low)no full probability distributionN/AGemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 69 eligible decisionssmall: N/Amedium stake: 6.9% (2 / 29), under 200 decisionsbig stake: 2.5% (1 / 40), under 200 decisionsQwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)Risk pays: 30 eligible decisionssmall: N/Amedium stake: 0.0% (0 / 14), under 200 decisionsbig stake: 0.0% (0 / 16), under 200 decisionsd1-ultraRisk pays: 606 eligible decisionssmall: N/Amedium stake: 59.6% (140 / 235)big stake: 68.5% (254 / 371)GLM-5.3 (low)Risk pays: 63 eligible decisionssmall: N/Amedium stake: 8.0% (2 / 25), under 200 decisionsbig stake: 2.6% (1 / 38), under 200 decisionsJev 1.14Risk pays: 152 eligible decisionssmall: N/Amedium stake: 59.6% (34 / 57), under 200 decisionsbig stake: 53.7% (51 / 95), under 200 decisionsGLiDE (API)Risk pays: 548 eligible decisionssmall: N/Amedium stake: 54.9% (117 / 213)big stake: 58.5% (196 / 335)Qwen3.8 Flash (off)Risk pays: 70 eligible decisionssmall: N/Amedium stake: 37.1% (13 / 35), under 200 decisionsbig stake: 45.7% (16 / 35), under 200 decisionsDeepSeek-V4.1-Flash (off · API · 1017)Risk pays: 57 eligible decisionssmall: N/Amedium stake: 33.3% (6 / 18), under 200 decisionsbig stake: 12.8% (5 / 39), under 200 decisionsClefRisk pays: 155 eligible decisionssmall: N/Amedium stake: 89.6% (43 / 48), under 200 decisionsbig stake: 92.5% (99 / 107), under 200 decisionsGNUbg weakno full probability distributionN/A

Three bars per setup: small, medium, big stake (Risk pays only); each bar is the share of eligible decisions in that band that are close choices. small medium big Rows in the Confidence table's order, not ranked by this share. A big stake means playing safe costs more by the engine's value, not that the decision was harder.

0%25%50%75%100%GPT-6.1 Sol (high)no full probability distributionN/AGPT-6 Astra (medium)no full probability distributionN/AGPT-6 Astra (low)no full probability distributionN/AGPT-6.1 Sol (low)no full probability distributionN/AGemini 4 Argon (high)no full probability distributionN/AClaude Fable 5.1 (high)no full probability distributionN/AClaude Opus 5.5 (high)no full probability distributionN/AClaude Opus 5.5 (low)no full probability distributionN/AGemini 4 Argon (low)no full probability distributionN/AGrok 4.7 (high)no full probability distributionN/AKimi K3.1 Pro (off)no full probability distributionN/AGNUbg weakno full probability distributionN/A

Three bars per setup: small, medium, big stake (Risk pays only); each bar is the share of eligible decisions in that band that are close choices. small medium big Rows in the Confidence table's order, not ranked by this share. A big stake means playing safe costs more by the engine's value, not that the decision was harder.

0%25%50%75%100%GPT-6.1 Sol (high)no full probability distributionN/AGPT-6 Astra (medium)no full probability distributionN/AGPT-6 Astra (low)no full probability distributionN/AGPT-6.1 Sol (low)no full probability distributionN/AGemini 4 Argon (high)no full probability distributionN/AClaude Fable 5.1 (high)no full probability distributionN/AClaude Opus 5.5 (high)no full probability distributionN/AClaude Opus 5.5 (low)no full probability distributionN/AGemini 4 Argon (low)no full probability distributionN/AGrok 4.7 (high)no full probability distributionN/AKimi K3.1 Pro (off)no full probability distributionN/AGNUbg weakno full probability distributionN/A

Three bars per setup: small, medium, big stake (Risk pays only); each bar is the share of eligible decisions in that band that are close choices. small medium big Rows in the Confidence table's order, not ranked by this share. A big stake means playing safe costs more by the engine's value, not that the decision was harder.

Analysis pending Shared positions: Analysis pending.

Probabilities over plays, not sampled reasoning. A decision is a close choice when the second most likely play has at least half the probability of the first (internally: split). N/A for setups without a full probability distribution. Method

Every measure of every setup, one tab per measure; each value with its 95% interval, its count and denominator or its price label under it.

Risk pays vs safety pays

0%25%50%75%100%Jev 1.13decision model · APIrisks found 27.0%27%right safety 92.1%92%DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release)risks found 29.1%29%right safety 89.5%89%GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsrisks found 33.0%33%right safety 84.6%85%Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsrisks found 33.4%33%right safety 82.5%82%celeris-1-decisiondecision model · APIrisks found 23.9%24%right safety 91.7%92%Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)risks found 23.5%24%right safety 91.7%92%Cygnetdecision model · our GPUsrisks found 27.3%27%right safety 86.8%87%Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)decision model · our GPUsrisks found 33.1%33%right safety 80.7%81%SubQ Snap Previewdecision model · APIrisks found 37.9%38%right safety 75.1%75%GPT-6 Luna (Decisions API)decision model · APIrisks found 21.6%22%right safety 89.7%90%Winnow-12Bdecision model · our GPUsrisks found 33.7%34%right safety 76.6%77%Mercury Decidedecision model · APIrisks found 18.7%19%right safety 88.1%88%LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)risks found 31.5%31%right safety 73.4%73%Perplexity Decider Flash v1.2decision model · our GPUsrisks found 31.3%31%right safety 66.9%67%Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell)decision model · our GPUsrisks found 37.3%37%right safety 59.8%60%LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell)risks found 28.8%29%right safety 62.7%63%GNUbg weakreferencerisks found 66.6%67%right safety 67.6%68%Randomreferencerisks found 24.8%25%right safety 25.0%25%Safestreferencerisks found 0.0%0%right safety 100.0%100%

Risks found (Risk pays) Right safety (Safety pays)

0%25%50%75%100%DeepSeek-V4.1-Flash (low · API + NVIDIA RTX PRO 6000 Blackwell · first release)reasoning low · letter probabilities · APIrisks found 45.0%45%right safety 93.0%93%DeepSeek-V4.1-Flash (low · API · 1017)reasoning low · letter probabilities · APIrisks found 42.5%43%right safety 94.5%94%Claude Sonnet 5.5 (high)reasoning high · one letter · subscriptionrisks found 41.8%42%right safety 94.6%95%Claude Fable 5.1 (low)risks found 41.0%41%right safety 94.6%95%Grok 4.7 (low)risks found 42.3%42%right safety 93.0%93%Gemini 3.8 Flash (high)reasoning high · one letter · subscriptionrisks found 36.1%36%right safety 97.6%98%Gemini 3.8 Flash (low)reasoning low · one letter · subscriptionrisks found 33.8%34%right safety 96.5%96%Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning on · letter probabilities · our GPUsrisks found 31.9%32%right safety 97.1%97%d1-ultrarisks found 42.0%42%right safety 86.0%86%Claude Sonnet 5.5 (low)reasoning low · one letter · subscriptionrisks found 30.9%31%right safety 96.2%96%Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning off · letter probabilities · our GPUsrisks found 33.2%33%right safety 92.1%92%Qwen3.8-27B (low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark)reasoning low · letter probabilities · our GPUsrisks found 33.8%34%right safety 90.2%90%GLM-5.3 (low)reasoning low · letter probabilities · APIrisks found 30.7%31%right safety 92.7%93%GLiDE (API)risks found 36.9%37%right safety 84.0%84%Qwen3.8 Flash (off)reasoning off · letter probabilities · APIrisks found 34.4%34%right safety 84.6%85%Jev 1.14risks found 29.1%29%right safety 89.4%89%DeepSeek-V4.1-Flash (off · API · 1017)reasoning off · letter probabilities · APIrisks found 25.4%25%right safety 91.7%92%Clefrisks found 39.5%39%right safety 75.6%76%GNUbg weakreferencerisks found 62.7%63%right safety 61.9%62%Random (computed)reference, computedrisks found 25.1%25%right safety 28.8%29%Safest (computed)reference, computedrisks found 0.0%0%right safety 100.0%100%

Risks found (Risk pays) Right safety (Safety pays)

0%25%50%75%100%GPT-6 Astra (low)reasoning low · one letter · subscriptionrisks found 57.2%57%right safety 94.4%94%GPT-6 Astra (medium)reasoning medium · one letter · subscriptionrisks found 54.9%55%right safety 94.4%94%Gemini 4 Argon (high)reasoning high · one letter · subscriptionrisks found 50.8%51%right safety 97.4%97%GPT-6.1 Sol (low)risks found 51.7%52%right safety 93.8%94%Claude Fable 5.1 (high)reasoning high · one letter · subscriptionrisks found 50.7%51%right safety 94.0%94%GPT-6.1 Sol (high)risks found 50.6%51%right safety 93.7%94%Claude Opus 5.5 (high)reasoning high · one letter · subscriptionrisks found 43.4%43%right safety 96.6%97%Gemini 4 Argon (low)reasoning low · one letter · subscriptionrisks found 47.8%48%right safety 90.7%91%Grok 4.7 (high)reasoning high · one letter · subscriptionrisks found 44.0%44%right safety 93.0%93%Claude Opus 5.5 (low)reasoning low · one letter · subscriptionrisks found 40.1%40%right safety 93.8%94%Kimi K3.1 Pro (off)reasoning off · one letter · APIrisks found 42.4%42%right safety 89.5%89%GNUbg weakreferencerisks found 58.8%59%right safety 65.5%65%Random (computed)reference, computedrisks found 25.1%25%right safety 30.1%30%Safest (computed)reference, computedrisks found 0.0%0%right safety 100.0%100%

Risks found (Risk pays) Right safety (Safety pays)

Hit risk only: being hit. Gammon risk and priming are not in it. Setups played different positions; league comparisons are not paired.

Can a model predict where it will waver?

Close choices: does a setup know where it hesitates?

Scope. Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark was started but not completed: two reboots and a GPU fault interrupted it, and it was stopped after 311 of 500 play calls on 2026-10-25. Its partial data is not reported.

Each team answered the same 100 positions 5 times, and once more in a separate call naming the features it expects to differ between the plays it might choose (hitting risk, making a point, the race, the prime). Observed set: the features that differ between any of the 5 chosen plays. Primary score: exact agreement of the predicted set with the observed set. Secondary score, declared secondary: partial overlap of the two sets. Baseline: the team's own predictions shuffled within phase, 1,000 seeded shuffles; p = the share of shuffles scoring at least the observed mean. Stage B asks again and looks at repeated answers; it is separate from stage A (Confidence tab), which reads recorded league decisions.

Swipe for more columns →

Stage B: five teams: per team, against its own shuffled baseline
Team and settingTemperaturePositions scoredExact agreement (primary score)Shuffled baseline, mean [2.5–97.5%]pExact agreement against its baselinePartial overlap (secondary score)Shuffled baseline, mean [2.5–97.5%]pPartial overlap against its baselineSame play in all 5 answersPlay changed, features identicalPositions where a feature changedFeatures predicted per position, meanLeft out, by reason
Claude Opus 5.5higheffort high (default, adaptive thinking)known: 1.0none sent; 1.0 is Anthropic's documented default1000.340.275 [0.23, 0.32]0.004above the baseline range0.4020.315 [0.268, 0.365]0.000above the baseline range7110191.19none
GPT-6 Astramediumreasoning effort medium (the model's default)not disclosednone sent; not disclosed1000.200.176 [0.15, 0.20]0.054inside the baseline range: not clearly above it0.2130.189 [0.160, 0.212]0.023above the baseline range90551.72none
GPT-6.1 Solhighreasoning effort highnot disclosednone sent; not disclosed1000.200.139 [0.11, 0.17]0.000above the baseline range0.2230.163 [0.127, 0.193]0.000above the baseline range8010101.64none
GPT-6.1 Sollowreasoning effort lownot disclosednone sent; not disclosed1000.130.111 [0.08, 0.14]0.154inside the baseline range: not clearly above it0.1690.135 [0.102, 0.165]0.010above the baseline range8010101.75none
Gemini 3.8 Flashlowthinking level lowknown: 1.01.0, sent1000.060.018 [0.01, 0.04]0.000above the baseline range0.0650.037 [0.024, 0.050]0.000above the baseline range92262.12none

Not a ranking: each team is compared with its own shuffled baseline, never with another team. Rows are in fixed alphabetical order and the table cannot be sorted.

A change, or no change, across five answers doesn't measure the model's whole behaviour.

Against its baseline: above the range = the observed mean is above the baseline's 97.5% point. Inside the range = not clearly above the baseline; a result that is not significant is not evidence that the team equals its baseline.

Most positions have no changed feature: the five answers chose the same play (71–92 positions per team), or changed play without changing a feature. So much of each score comes from predicting “none” where the team does not hesitate.

No team gave the same play in all five answers on every scored position.

Every team kept its league settings. Temperature: the value sent or the provider's documented default where known, otherwise not disclosed.

Exact agreement against the shuffled baseline

0.000.100.200.300.40Claude Opus 5.5 (high): exact agreement 0.34; shuffled baseline 0.275 [0.23, 0.32]; p 0.004; above the baseline rangeClaude Opus 5.5 (high)p 0.004, above the range0.34GPT-6 Astra (medium): exact agreement 0.20; shuffled baseline 0.176 [0.15, 0.20]; p 0.054; inside the baseline range: not clearly above itGPT-6 Astra (medium)p 0.054, inside the range0.20GPT-6.1 Sol (high): exact agreement 0.20; shuffled baseline 0.139 [0.11, 0.17]; p 0.000; above the baseline rangeGPT-6.1 Sol (high)p 0.000, above the range0.20GPT-6.1 Sol (low): exact agreement 0.13; shuffled baseline 0.111 [0.08, 0.14]; p 0.154; inside the baseline range: not clearly above itGPT-6.1 Sol (low)p 0.154, inside the range0.13Gemini 3.8 Flash (low): exact agreement 0.06; shuffled baseline 0.018 [0.01, 0.04]; p 0.000; above the baseline rangeGemini 3.8 Flash (low)p 0.000, above the range0.06

shuffled baseline, 2.5–97.5% baseline mean exact agreement, above the range inside the range: not clearly above Each team against its own baseline; not a ranking.

Data: close-choices-b.json · Method (Method spec 2026-10-24 §9.9-9.13)

Scorer
GNU Backgammon 1.08, 2-ply, on every legal play of every decision
Decisions scored
92,888
Leagues
League 1: decision models and models on our own hardware; League 2: workhorse chat models, closed and open; League 3: frontier chat models on subscriptions
Results
f24d1f41, data as of 16 November 2026, 21:24 UTC
Published
Every result per setup (results.json). Not published: the positions, the plays chosen in them and the engine's best play for each (why). No internal address, key or request header.

What this measures. Risk calls per setup: was each risk the right one, and was safe right when it played safe? Risk judgement: 100 = every call right, 50 = always safe. Error rate: how much equity each move gives away, judged by the engine, not by wins. Hit risk only: being hit. Gammon risk and priming are not in it. Setups met different positions; risk judgement is not adjusted for how hard they were. Paired comparisons wait for the shared positions.

About these tables

One table per league. League tables compare only the players inside one league; they are never merged into one overall score. A running league's numbers are Provisional and change as matches finish. Click a row to open the setup, or a match to open its decisions.

Earlier version: v1 leagues and teams Earlier version

Earlier version: plays were listed in the runner's order, which tends to put the best play early. When Gemini 4 Argon answered the same 1,000 positions with the plays shuffled, its error rate changed by +1.9 (95% interval −0.2 to +4.2), so the effect is not proven.

Earlier version The sections below are kept from the earlier report, unchanged. Their numbers come from the earlier version and are never ranked with current numbers. The prompt and how it was built are on the Method page.

The setup

Earlier version Plays were listed in the runner's order, which tends to put the best play early.

Each model plays every other model once per league (a single round robin). A match is played in rounds of two games with the same dice: in the second game the players swap seats and the same sequence of rolls is repeated, so both face the same luck. Games score 1 point, 2 for a gammon and 3 for a backgammon; there is no doubling cube.

  • Match length: 3 rounds, or fewer: the first player to 5 points after a full round (both games) wins if ahead. A tie gets one extra round, then counts as a draw.
  • League table: 3 points for a match win, 1 for a draw; ties broken by more games won, then fewer games lost.
  • Turn limit: a game still running after 300 turns (both sides counted) goes to the lower pip count for 1 point.
  • Clock: 30 seconds free per decision, then a 5-minute bank per game. Running out loses the game.
  • Legality: the program generates every legal play and the model chooses one of them, so a model can never make an illegal move. A roll with only one legal play is played automatically.
  • Memory: every decision is a separate request. Models are not told how earlier games went.

With seven models, each league has 21 matches and one model sits out each matchday. A match starts as soon as both of its models are free, and no model plays two matches at once.

Models

Earlier version Plays were listed in the runner's order, which tends to put the best play early.

All seven answer the same kind of request: a text state and one choice question whose options are the legal plays. Two are hosted services; the other five run on our own GPUs.

Jev 1.13 · TypeSafe, hosted
TypeSafe's decision model, served through its System One API; the request format the others copy.
GPT-6 Luna · OpenAI Decisions API, hosted
OpenAI's decision endpoint with gpt-6-luna. The only one that sometimes declines to choose; that happened 3 times in early leagues, on positions with 97 to 439 legal plays, and never in the current format.
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · Perplexity, self-hosted
The released pplx-decider-v1.1-27b checkpoint, run with its own server code. It accepts up to 8,192 tokens, so long play lists are sent in groups of 25 (see below). It replaced Perplexity's hosted Decider 27B from league #6 on.
Clef · Cloudflare, self-hosted
A 27B decision model that scores all options in one pass. Groups of up to 110 options.
Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell · open reimplementation of Jev, self-hosted
A LoRA adapter and decision head on a 27B base model. It scores every option as its own sequence that repeats the whole question, with a 4,096-token limit per sequence, which is what limits the prompt's length for every model.
Cygnet · blockbrain-ai recipe, self-hosted
Google's gemma-4-12B-it, unchanged, behind a server that labels the options A, B, C… and reads the model's probability for each letter. Groups of up to 26 options.
Gemma 4 31B · our configuration of the Cygnet recipe
The same letter-reading server in front of gemma-4-31B-it, served in FP8. Not a trained decision model: a general chat model whose answer-letter probabilities are read directly. Its calibration setting was fitted for Cygnet's 12B model; that changes its reported confidence, not which play it picks.

When a position has more legal plays than a model's group size, the plays are split into equal groups, the model picks a winner in each, then chooses among the winners. Liquid's d1 played in the early leagues and was dropped after its free tier rate-limited a league to a stop.

Error rate

Earlier version Plays were listed in the runner's order, which tends to put the best play early.

Wins and losses depend heavily on the dice. A backgammon engine can instead judge every decision directly: for each move it finds the best play, and measures how much winning chance, in equity (expected points per game), the chosen play gave away. Averaged over all decisions and multiplied by 500, this is the error rate (performance rating, PR) that backgammon players use, here for checker plays only; forced moves are not counted. eXtreme Gammon's human scale: under 2.5 world champion, 2.5 to 5 world class, 5 to 7.5 expert, 7.5 to 12.5 advanced, 12.5 to 17.5 intermediate, 17.5 to 22.5 casual, 22.5 to 30 beginner, 30 or more distracted. Human ratings also count doubling cube decisions, which these games do not have, so the comparison is approximate.

We scored all 104,045 decisions the models made in every league with GNU Backgammon 1.08. Every legal play was evaluated at 0-ply, then the five best and the chosen play again at 2-ply; the loss is the 2-ply equity of the best play minus that of the chosen one. As a check, on the opening roll the plays that the rollout table calls best lose 0.0002 equity on average, other plays 0.063.

By model, leagues #8, #11 and #12

ModelDecisionsError rateEngine's best playBlunders#8#11#12#13
Gemma 4 31B1,89425.4 ± 2.454%19.5%22.726.327.228.8
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark1,94232.7 ± 2.949%24.5%33.033.131.934.2
Jev 1.132,06635.7 ± 2.946%26.7%34.236.236.732.0
Clef2,10035.8 ± 3.047%25.6%35.740.530.642.8
GPT-6 Luna2,04439.4 ± 3.546%25.6%40.135.941.737.5
Cygnet2,07843.2 ± 3.442%30.2%46.945.636.740.6
Open-Jev 27B1,75859.6 ± 4.636%36.5%65.863.149.558.1

Error rate: mean equity lost per decision × 500, with a 95% interval that treats decisions as independent (moves within a game are not, so the true interval is somewhat wider). "Engine's best play": share of decisions where the chosen play lost less than 0.001. "Blunders": decisions losing more than 0.08 equity. The last four columns are the error rate in each league; #13 adds the move history.

  • The engine ranks the models almost as the league tables do. Gemma 4 31B is clearly best and Open-Jev 27B clearly worst. Decider v1.1, Jev, Clef and GPT-6 Luna are within about one interval of each other, which is why their league places move around; Cygnet is a little behind them.
  • Every model blunders often: from one move in five (Gemma 4 31B) to one in three (Open-Jev 27B). Even Gemma 4 31B chooses the engine's best play only about half the time.
  • The prompt mattered more than the model. Before move consequences were computed, the whole field averaged 120 to 138; after, 36 to 40. The worst model now is better than the average model was then.

Results

Earlier version Plays were listed in the runner's order, which tends to put the best play early.

The standings below use the three finished leagues with computed move consequences and no move history: #8, #11 and #12 (7 models, same prompt, different dice). League #13, with move history, finished 1. Gemma 4 31B, 2. Jev 1.13, 3. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark, 4. Cygnet, 5. GPT-6 Luna, 6. Clef, 7. Open-Jev 27B.

Final places

Model1st2nd3rd4th5th6th7th
Gemma 4 31B2100000
Jev 1.131001100
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark0210000
Clef0011010
Cygnet0011010
GPT-6 Luna0000300
Open-Jev 27B0000003

In league #12, Jev 1.13 and GPT-6 Luna finished level on every tiebreak and share 5th place.

Across the three leagues

ModelMatches W-D-LGames won-lostPoints for-againstGammons and backgammons wonPicked the safest playShots left on averageTime per decision
Gemma 4 31B14-1-367-2791-362384%5.10.30 s
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark12-0-656-4079-562077%6.11.14 s
Jev 1.1311-0-753-4971-611682%5.10.26 s
Clef9-1-853-4581-642168%7.01.13 s
Cygnet8-0-1052-5465-721182%4.60.22 s
GPT-6 Luna8-0-1048-5462-681187%3.80.20 s
Open-Jev 27B0-0-1815-7518-110265%8.51.69 s

"Safest play": the option whose position the opponent hits with the fewest of its 36 rolls, read from the numbers the model was shown. "Shots left": opponent rolls out of 36 that hit after the chosen play. Both are measured on decisions sent as a single request (about 1,700 to 2,100 per model); the safest available play left 2.7 to 4.2 shots on average, a typical option about 13.

What the numbers say

  • Gemma 4 31B is the strongest player here, and it is not a trained decision model: a general chat model whose answer-letter probabilities are read. It lost 3 of 18 matches and won the most gammons.
  • Safe is not the same as good. GPT-6 Luna picks the safest play most often and leaves the fewest shots, yet finished 5th in all three leagues. Winning also takes making points and building a prime, which costs some risk; Gemma and Clef take more risk and win more gammons.
  • Open-Jev 27B collapsed in this format. It finished 2nd in league #3, where options were just notations, and last in every league since, with the most shots left and the highest error rate. We have not found out why; the numbers it is shown are the same as everyone else's.
  • Speed does not decide it. The three fastest models are spread across the table, and no model came close to its clock.

Can a team beat the best model?

Earlier version Plays were listed in the runner's order, which tends to put the best play early.

Every model reports a probability for each option. If those probabilities say when a model is right, a team could ask all seven and play the answer of whichever is most sure. To test this, we sent the same 1,000 recorded positions (2 to 25 legal plays each, drawn from leagues #8, #11, #12 and #13) to all seven models with exactly the request the original player received, and scored every legal play with GNU Backgammon at 2-ply: 7,000 answers and 9,439 evaluated plays. Any team and any rule can then be scored offline on the same positions.

Confidence says how hard a position is, not who is right

Within each model, confidence does carry information: across its league decisions, a model is right more often when it is more confident (AUROC 0.66 to 0.74, where 0.5 is chance). But the models use very different scales:

ModelEngine's best playAverage stated probabilityWhen it is most sureit is right
Gemma 4 31B57%0.85 overconfident0.8 – 1.065%
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark51%0.580.8 – 1.084%
Jev 1.1347%0.550.8 – 1.081%
Clef46%0.28 underconfident0.6 – 0.889%
GPT-6 Luna46%0.560.8 – 1.065%
Cygnet44%0.67 overconfident0.8 – 1.063%
Open-Jev 27B36%0.21 underconfident0.6 – 0.892%

From 17,579 league decisions sent as a single request; "right" means the engine's best play (loss below 0.001 equity). The stated probability is the one the model gave its chosen play; the models' own "confidence" fields are defined differently from one another, so they are not used.

Clef and Open-Jev look like the most reliable models when they are sure, but they are only sure on easy positions, where everyone else is right as well. The test that matters is what happens when models disagree on the same position:

When Gemma 4 31B and this model disagreeGemma has the better playThe more confident model has the better play
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark62%62%
Clef64%64%
Jev 1.1369%67%
GPT-6 Luna73%70%
Cygnet77%75%
Open-Jev 27B76%76%

Following the more confident model is never better than simply following Gemma 4 31B.

Teams

Team and ruleError rateEngine's best playDifference from Gemma alone (95%)
Gemma 4 31B alone23.554.4%–
All 7: highest stated probability24.053.8%+0.4 (−0.2 to +1.2)
All 7: lowest expected loss, calibrated26.052.9%+2.5 (+0.7 to +4.3)
All 7: most likely correct, calibrated27.751.9%+4.2 (+2.2 to +6.4)
All 7: probabilities added up27.752.0%+4.2 (+2.1 to +6.6)
Best of 120 formations: Clef + Decider v1.1 + Gemma, probabilities added up23.454.8%−0.2 (−0.7 to +0.2)
Oracle: the best answer any of the 7 gave12.770.1%−10.8

Error rate on the 1,000 shared positions (lower is better). Calibrated rules map each model's stated probability to its observed accuracy or loss with isotonic regression, fitted on one half of the positions and scored on the other (two-fold cross-fitting). Intervals are bootstrap intervals over positions, paired with Gemma 4 31B on the same positions. The best formation was chosen on the same positions it is scored on, which flatters it.

  • No team beats the best single model. Raw confidence ties with Gemma 4 31B only because its inflated probabilities win almost every vote; correcting the scales makes the team worse.
  • The room for a team is large. Gemma 4 31B misses the engine's best play on 456 of the 1,000 positions, and on a third of those another model found it. An oracle would halve the error rate. What is missing is a way to tell which answer is the right one.
  • The data is kept for further formations. Other selectors (agreement between models, the number of legal plays, the stage of the game) can be tested on the same 1,000 positions without new requests.

Speed and cost

Earlier version Plays were listed in the runner's order, which tends to put the best play early.

Measured on the same 1,000 recorded positions as the team test, with exactly the request each original player received. Besides the seven league players, the table includes general language models we tried the same way: DeepSeek-V4.1-Flash, Qwen3.8 Flash Next and GLM-4.7-Flash are read through their probabilities for each answer letter, like Gemma 4 31B. GPT-6 Astra, Claude Opus 5.5, Gemini 4 Argon and Gemini 3.8 Flash return no such probabilities: each answers one letter, constrained to the option letters, at low reasoning effort (Gemini's lowest available level). Gemini 4 Argon reported no reasoning tokens at that level, so it answered directly. DeepSeek-V4.1-Flash also ran a second time with reasoning on (lowest effort), answering one constrained letter after reasoning. GPT-6 Astra and Claude Opus 5.5 ran through ChatGPT and Claude subscriptions.

ModelRuns onError rateTime per decision
median / 90th pct
In flightEstimated API cost
per 1,000 decisions
GPT-6 AstraChatGPT subscription11.24.60 s / 10.95 s4$36.45 API-equivalent
Gemini 4 ArgonGoogle Gemini API16.94.09 s / 28.22 s4$6.97
Claude Opus 5.5Claude subscription17.33.43 s / 6.66 s4$21.65 API-equivalent
DeepSeek-V4.1-Flash, reasoning onour GPUs: 418.1106 s / 288 s8own hardware
Gemini 3.8 FlashGoogle Gemini API18.32.87 s / 7.96 s4$3.25
Gemma 4 31Bour GPUs: 123.50.59 s / 0.87 s4own hardware
DeepSeek-V4.1-Flashour GPUs: 427.50.57 s / 0.98 s4own hardware
Qwen3.8 Flash Nextour GPUs: 431.10.46 s / 0.71 s4own hardware
Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Sparkour GPUs: shares 231.80.82 s / 1.35 s1own hardware
Jev 1.13TypeSafe API33.60.25 s / 0.29 s4$0.16
Clefour GPUs: 134.10.94 s / 1.28 s1own hardware
GPT-6 LunaOpenAI Decisions API35.00.22 s / 0.27 s4$0.32
Cygnetour GPUs: 138.90.78 s / 1.12 s4own hardware
GLM-4.7-Flashour GPUs: 150.60.20 s / 0.27 s4own hardware
Open-Jev 27Bour GPUs: 159.91.38 s / 1.91 s1own hardware

Error rate on the 1,000 shared positions (lower is better). Time is wall-clock per request, from sending it to receiving the answer, with the stated number of requests in flight. Our GPUs are NVIDIA RTX PRO 6000 Blackwell cards with 96 GB each, in one machine. The seven league players were asked together, so the self-hosted ones shared the machine; the other models were each asked alone. A run of GLM-5.3-Flash is left out: it played close to random, most likely because of how we had to run it on this machine.

The plays are listed in the order the runner generated them, which tends to put the best play early. Asked again with the same plays in a random order, Gemini 4 Argon's error rate rose from 16.9 to 18.9 (difference 1.9, 95% interval −0.3 to 4.1): probably a small help from the order, not yet a proven one. The next version of the benchmark shuffles the plays for every model.

Cost basis: the provider's list price × the tokens its API reported; the bill itself was not checked. Jev 1.13 ($0.042 per 1M input tokens, output free) and GPT-6 Luna on OpenAI's Decisions API ($0.10 per 1M input tokens, output free): prices of 8 October 2026, tokens averaged over their league decisions with 2 to 25 legal plays (3,740 and 3,238 input tokens per decision). Gemini 3.8 Flash ($0.75 per 1M input tokens, $3.75 per 1M output tokens, reasoning billed as output; introductory price of 8 October 2026, which doubles on 1 January 2027): tokens measured on 100 of the positions, 3,373 input and 192 output per decision on average. Gemini 4 Argon ($2 per 1M input, $10 per 1M output; price until 30 November 2026, doubling on 1 December): tokens reported for all 1,000 positions, 3,462 input and 5 output per decision. GPT-6 Astra ($10 per 1M input, $50 per 1M output) and Claude Opus 5.5 ($4 input, $20 output): API list prices of 8 October 2026 × the tokens reported for all 1,000 positions (3,213 input and 86 output, and 4,424 input and 198 output per decision). Both ran on subscriptions, which are not billed per token, so these are API-equivalent estimates. Self-hosted models have no per-request price; a 7-model league of about 690 decisions per player took about 30 minutes on the four GPUs. DeepSeek with reasoning used 6,657 output tokens per decision on average.

  • The cheapest strong player is one we run ourselves. Gemma 4 31B costs only GPU time, needs one card, and answers in about 0.6 s.
  • The best players cost most and are slowest. GPT-6 Astra makes the fewest errors by far, but at API prices it costs about a hundred times as much per decision as GPT-6 Luna, and one decision in ten takes 11 seconds or more. Gemini 4 Argon and Claude Opus 5.5 show no clear difference (16.9 against 17.3; difference 0.3, 95% interval −1.9 to 2.5), with Argon about a third of Opus's price; Gemini 3.8 Flash (18.3) is close behind at half Argon's price. Even asking Astra, Opus and Gemini 3.8 Flash together and taking the majority answer is worse than Astra alone (14.4).
  • Reasoning helps an open model a lot, at a steep time cost. With reasoning on, DeepSeek V4.1 Flash's error rate falls from 27.5 to 18.1 (95% interval of the gain 6.1 to 12.7), close to Opus's, but a decision takes about 100 seconds instead of half a second.
  • The hosted decision services are fastest and cheapest per decision, at about 0.2 s and well under a dollar per 1,000 decisions, but they play worse than Gemma 4 31B.

Limits

Earlier version Plays were listed in the runner's order, which tends to put the best play early.

  • Small sample. Three finished leagues of 21 matches in the main format, about 340 games. Placings in the middle of the table move from league to league; the error rate is the steadier measure.
  • The engine evaluation is not the strongest available. Moves are judged at 2-ply with GNU Backgammon, not by full rollouts or eXtreme Gammon, and in cubeless money equity: the games have no doubling cube, and match score is ignored. Small differences between close plays can be wrong, but the losses that count as blunders and the gaps between models are much larger than that.
  • Models only play each other. Their absolute strength against human or engine players is unknown. On the engine's scale none of them reaches expert level; the best, GPT-6 Astra, plays like an advanced human.
  • Configuration matters. Five models run on servers set up by us, and Gemma 4 31B uses a recipe built for a different model. Hosted services can change without notice.
  • The prompt steers. The question and the strategy text say what a good move looks like, so the results measure how well models follow that guidance with the numbers in front of them, as much as how well they play.

The page as it was before this site went live, with its final league data: archive, 23 October 2026.