Risk calls

92,888 decisions scored in 3 leagues · GNU Backgammon 1.08, 2-ply

All models

Models
Class
Rows
Measure
100 = every call right, 50 = always safe; higher is better.Did the model improve after lessons? Yes: the change in risks found is above the threshold and right safety falls by no more than 5 pp. No: either fails. Shown once every run is complete; before that, the model's state (running, queued, waiting and why, not tested and the rule).500 × mean 2-ply equity loss per decision, the checker-play performance rating (PR) backgammon players use; lower is better.Dollars per 1,000 decisions at each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid. Lower is better.

Risk judgement all leagues · 95% interval · higher is better

View
  1. 1GPT-6 Astra high · chat · late entry74.4
  2. GPT-6 Astra low · chat75.8
  3. GPT-6 Astra medium · chat74.7
  4. 2Gemini 4 Argon high · chat74.1
  5. Gemini 4 Argon low · chat69.2
  6. 3Claude Fable 5.1 high · chat72.3
  7. Claude Fable 5.1 low · chat67.8
  8. 4GPT-6.1 Sol high · chat72.2
  9. GPT-6.1 Sol low · chat72.7
  10. 5Claude Opus 5.5 high · chat70.0
  11. Claude Opus 5.5 low · chat66.9
  12. 6DeepSeek-V4.1-Flash low · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat69.0
  13. DeepSeek-V4.1-Flash low · API · 1017 · chat68.5
  14. DeepSeek-V4.1-Flash off · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat59.3
  15. DeepSeek-V4.1-Flash off · API · 1017 · chat58.5
  16. 7Grok 4.7 high · chat68.5
  17. Grok 4.7 low · chat67.7
  18. 8Claude Sonnet 5.5 high · chat68.2
  19. Claude Sonnet 5.5 low · chat63.5
  20. 9Gemini 3.8 Flash high · chat66.9
  21. Gemini 3.8 Flash low · chat65.1
  22. 10Muse Ember max · chat · late entry66.0
  23. 11Kimi K3.1 Pro off · chat65.9
  24. 12Gemini Decision 1 20261111 · decision · late entry64.7
  25. 13Gemma 4 31B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat64.5
  26. Gemma 4 31B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat62.6
  27. 14d1-ultra decision64.0
  28. 15Mammoth-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry63.2
  29. 16Drex Max NVIDIA RTX PRO 6000 Blackwell · decision · late entry62.9
  30. 17Sieve 31B v3 NVIDIA DGX Spark (GB10) · decision · late entry62.7
  31. 18Qwen3.8-27B low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat62.0
  32. Qwen3.8-27B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat57.6
  33. 19GLM-5.3 low · chat61.7
  34. 20KnowLine-31B-Gen1.5 NVIDIA DGX Spark (GB10) · decision · late entry61.6
  35. 21REAX 1.3 NVIDIA RTX PRO 6000 Blackwell · decision · late entry60.5
  36. 22GLiDE API · decision60.4
  37. 23Turion Jev 31B NVIDIA DGX Spark (GB10) · decision · late entry59.6
  38. 24Jev 1.13 decision59.5
  39. 25Qwen3.8 Flash off · chat59.5
  40. 26Jev 1.14 decision59.3
  41. 27Drex Max v1.1 decision · late entry59.1
  42. 28GLiDE v1.2 · NVIDIA RTX PRO 6000 Blackwell · decision58.8
  43. 29Drex 2.0 API · decision · late entry58.1
  44. Drex 2.0 NVIDIA RTX PRO 6000 Blackwell · decision · late entry53.0
  45. 30Perplexity Decider 27B v1.2 NVIDIA RTX PRO 6000 Blackwell · decision57.9
  46. 31celeris-1-decision decision57.8
  47. 32Elephant-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry57.7
  48. 33Clef NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision57.5
  49. 34StartLux-Decision-v1.1-4B-Instant NVIDIA DGX Spark (GB10) · decision · late entry57.2
  50. 35Cygnet NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision57.1
  51. 36Perplexity Decider 27B v1.1 NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision56.9
  52. 37SubQ Snap 1.1 decision · late entry56.7
  53. 38SubQ Snap Preview decision56.5
  54. 39Cerno 9B v1.2 NVIDIA DGX Spark (GB10) · decision · late entry56.1
  55. 40GPT-6 Luna Decisions API · decision55.7
  56. 41d1-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry55.2
  57. 42Winnow-12B NVIDIA RTX PRO 6000 Blackwell · decision55.2
  58. 43StartLux-Decision-v1.1-2B-Instant NVIDIA DGX Spark (GB10) · decision · late entry53.8
  59. 44Mercury Decide decision53.4
  60. 45d1-Omni-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry53.2
  61. 46LFM2.5-24B-A2B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat52.4
  62. LFM2.5-24B-A2B off · NVIDIA RTX PRO 6000 Blackwell · chat45.7
  63. 47KnowLine-2B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry51.1
  64. 48KnowLine 4B Gen5.5 NVIDIA DGX Spark (GB10) · decision · late entry49.7
  65. 49Perplexity Decider Flash v1.2 NVIDIA RTX PRO 6000 Blackwell · decision49.1
  66. 50Open-Jev-27B-v1.1 NVIDIA RTX PRO 6000 Blackwell · decision48.5
  67. 51GLiDE 2B NVIDIA DGX Spark (GB10) · decision · late entry46.7
  68. 52KnowLine-0.8B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry45.2
  69. 53StartLux-Decision-v1.1-0.8B-Instant NVIDIA DGX Spark (GB10) · decision · late entry44.6
  70. 54OneJev1.6-0.8B NVIDIA DGX Spark (GB10) · decision · late entry42.5
  71. 55GLiDE 0.8B NVIDIA DGX Spark (GB10) · decision · late entry41.7
  72. 56GLiDE 0.4B NVIDIA DGX Spark (GB10) · decision · late entry28.9
  73. 57GLiClass v4 Edge NVIDIA DGX Spark (GB10) · decision · late entry21.5
  74. GNUbg weak program67.1
  75. GNUbg weak program62.3
  76. GNUbg weak program62.1
  77. Safest program50.0
  78. Safest (computed) program50.0
  79. Safest (computed) program50.0
  80. Random program24.9
  81. Random (computed) program27.0
  82. Random (computed) program27.6

No model of this kind on the leaderboard.

  • OpenAI
  • Google
  • Anthropic
  • DeepSeek
  • xAI
  • Meta
  • Moonshot
  • Liquid AI
  • Base Labs
  • Nace.AI
  • Sthānika AI
  • Alibaba
  • Z.ai
  • PelaAI
  • REAX
  • Fastino
  • Turion.AI
  • TypeSafe
  • Perplexity
  • Celeris
  • Cloudflare
  • StartLux
  • Blockbrain Labs
  • Subquadratic
  • Apertix AI
  • EldanRing
  • Inception
  • Zefan Cai
  • OmniJev
  • Knowledgator
  • program
  • 50 = always safe
Counts 1, 2, 3 down the rows shown, by the measure chosen; no tiers.The model and its setting. Tap or point at a name for its details.100 = every call right, 50 = always safe; higher is better.500 × mean 2-ply equity loss per decision, the checker-play performance rating (PR) backgammon players use; lower is better.Dollars per 1,000 decisions at each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid. Lower is better.Median request seconds per decision (waits between retries not included). Lower is better.
1GPT-6 Astra high · chat · late entry Closed · FlagshipDetails74.471.8–77.0 (50 = always safe)8.8$59.356$59,356List price9.83
GPT-6 Astra low · chat Closed · FlagshipDetails75.873.1–78.5 (50 = always safe)11.8$49.737$49,737List price3.94
GPT-6 Astra medium · chat Closed · FlagshipDetails74.771.9–77.3 (50 = always safe)11.1$50.268$50,268List price4.40
2Gemini 4 Argon high · chat Closed · FlagshipDetails74.172.0–76.1 (50 = always safe)12.6$64.401$64,401List price27.5
Gemini 4 Argon low · chat Closed · FlagshipDetails69.266.2–72.3 (50 = always safe)16.5$10.204$10,204List price3.82
3Claude Fable 5.1 high · chat Closed · FlagshipDetails72.369.8–74.7 (50 = always safe)14.0$136.268$136,268List price12.7
Claude Fable 5.1 low · chat Closed · FlagshipDetails67.865.6–69.9 (50 = always safe)17.9$62.613$62,613List price1.54
4GPT-6.1 Sol high · chat Closed · WorkhorseDetails72.269.5–74.9 (50 = always safe)9.9$15.573$15,573List price13.1
GPT-6.1 Sol low · chat Closed · WorkhorseDetails72.769.9–75.5 (50 = always safe)12.1$10.228$10,228List price4.33
5Claude Opus 5.5 high · chat Closed · FlagshipDetails70.067.3–72.6 (50 = always safe)14.4$36.400$36,400List price6.35
Claude Opus 5.5 low · chat Closed · FlagshipDetails66.964.7–69.2 (50 = always safe)15.4$29.062$29,062List price3.92
6DeepSeek-V4.1-Flash low · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetails69.067.3–70.9 (50 = always safe)19.0$47.092$47,092GPU estimate43.8
DeepSeek-V4.1-Flash low · API · 1017 · chat Open · 552B MoE · 16B activeDetails68.566.4–70.6 (50 = always safe)17.1$4.869$4,869List price22.3
DeepSeek-V4.1-Flash off · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetails59.357.1–61.5 (50 = always safe)35.6$0.728$728GPU estimate0.48
DeepSeek-V4.1-Flash off · API · 1017 · chat Open · 552B MoE · 16B activeDetails58.556.6–60.5 (50 = always safe)35.2$0.244$244List price0.84
7Grok 4.7 high · chat Closed · WorkhorseDetails68.566.1–71.1 (50 = always safe)18.8$117.379$117,379List price122.3
Grok 4.7 low · chat Closed · WorkhorseDetails67.766.0–69.5 (50 = always safe)19.1$32.813$32,813List price24.3
8Claude Sonnet 5.5 high · chat Closed · WorkhorseDetails68.266.3–70.0 (50 = always safe)16.7$19.847$19,847List price6.63
Claude Sonnet 5.5 low · chat Closed · WorkhorseDetails63.561.6–65.4 (50 = always safe)20.1$13.541$13,541List price1.51
9Gemini 3.8 Flash high · chat Closed · WorkhorseDetails66.965.2–68.5 (50 = always safe)14.9$22.352$22,352List price9.59
Gemini 3.8 Flash low · chat Closed · WorkhorseDetails65.163.5–66.8 (50 = always safe)16.3$4.123$4,123List price2.16
10Muse Ember max · chat · late entry Closed · EfficientDetails66.063.9–68.2 (50 = always safe)20.1$6.863$6,863Billed15.3
11Kimi K3.1 Pro off · chat Open · 2.8T MoE · 104B activeDetails65.963.2–68.9 (50 = always safe)28.0$3.982$3,982List price2.11
12Gemini Decision 1 20261111 · decision · late entry Closed · EfficientDetails64.762.2–67.2 (50 = always safe)28.0$0.382$382Billed0.24
13Gemma 4 31B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetails64.562.9–66.2 (50 = always safe)19.6$34.943$34,943GPU estimate120.6
Gemma 4 31B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetails62.660.4–65.0 (50 = always safe)22.5$8.215$8,215GPU estimate0.31
14d1-ultra decision Closed · EfficientDetails64.061.7–66.3 (50 = always safe)30.3$0.145$145List price0.23
15Mammoth-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetails63.261.2–65.2 (50 = always safe)28.5$0.999$999GPU estimate0.72
16Drex Max NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetails62.961.1–64.7 (50 = always safe)30.4$1.285$1,285GPU estimate0.68
17Sieve 31B v3 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails62.760.6–64.9 (50 = always safe)27.5$0.794$794GPU estimate11.1
18Qwen3.8-27B low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetails62.060.1–63.9 (50 = always safe)29.0$39.023$39,023GPU estimate171.1
Qwen3.8-27B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetails57.656.1–59.0 (50 = always safe)45.0$4.551$4,551GPU estimate1.48
19GLM-5.3 low · chat Open · 753B MoEDetails61.759.7–63.8 (50 = always safe)30.9$2.917$2,917List price1.75
20KnowLine-31B-Gen1.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails61.659.6–63.5 (50 = always safe)26.2–3.45
21REAX 1.3 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 57BDetails60.558.7–62.4 (50 = always safe)27.8$1.545$1,545GPU estimate1.74
22GLiDE API · decision Closed · EfficientDetails60.458.3–62.6 (50 = always safe)33.2$0.153$153List price0.48
23Turion Jev 31B NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails59.657.7–61.4 (50 = always safe)29.7$0.293$293GPU estimate3.95
24Jev 1.13 decision Closed · EfficientDetails59.557.4–61.6 (50 = always safe)34.4$0.166$166List price0.22
25Qwen3.8 Flash off · chat Closed · EfficientDetails59.557.2–61.7 (50 = always safe)33.6$0.284$284List price1.70
26Jev 1.14 decision Closed · EfficientDetails59.357.1–61.4 (50 = always safe)31.7$0.173$173List price0.19
27Drex Max v1.1 decision · late entry Closed · Efficient (provisional)Details59.157.0–61.3 (50 = always safe)30.9$0.265$265List price0.44
28GLiDE v1.2 · NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails58.856.7–60.8 (50 = always safe)35.0$10.301$10,301GPU estimate1.41
29Drex 2.0 API · decision · late entry Open · 35B MoE · 3.0B activeDetails58.155.7–60.4 (50 = always safe)38.4$0.258$258Billed0.42
Drex 2.0 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 35B MoE · 3.0B activeDetails53.050.5–55.2 (50 = always safe)42.9$0.782$782GPU estimate0.40
30Perplexity Decider 27B v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails57.956.0–59.9 (50 = always safe)36.9$12.227$12,227GPU estimate0.94
31celeris-1-decision decision Closed · EfficientDetails57.855.9–59.7 (50 = always safe)35.6$0.150$150List price0.17
32Elephant-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 9.0BDetails57.755.8–59.6 (50 = always safe)44.6$0.685$685GPU estimate0.26
33Clef NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetails57.555.4–59.8 (50 = always safe)36.5$5.736$5,736GPU estimate1.13
34StartLux-Decision-v1.1-4B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetails57.255.0–59.3 (50 = always safe)40.1$0.072$71.58GPU estimate0.44
35Cygnet NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 12BDetails57.155.3–58.8 (50 = always safe)38.6$11.168$11,168GPU estimate0.37
36Perplexity Decider 27B v1.1 NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetails56.954.9–58.8 (50 = always safe)38.8$6.593$6,593GPU estimate1.97
37SubQ Snap 1.1 decision · late entry Closed · EfficientDetails56.754.8–58.6 (50 = always safe)40.2$0.150$150Billed0.23
38SubQ Snap Preview decision Closed · EfficientDetails56.553.8–59.1 (50 = always safe)43.9$0.193$193List price0.24
39Cerno 9B v1.2 NVIDIA DGX Spark (GB10) · decision · late entry Open · 9.0BDetails56.153.9–58.2 (50 = always safe)48.5$0.139$139GPU estimate1.96
40GPT-6 Luna Decisions API · decision Closed · EfficientDetails55.753.8–57.6 (50 = always safe)43.3$0.335$335List price0.23
41d1-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 1.9BDetails55.252.8–57.5 (50 = always safe)47.1$0.467$467GPU estimate0.06
42Winnow-12B NVIDIA RTX PRO 6000 Blackwell · decision Open · 12BDetails55.252.9–57.2 (50 = always safe)49.7$9.615$9,615GPU estimate2.46
43StartLux-Decision-v1.1-2B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetails53.851.3–56.6 (50 = always safe)50.0$0.052$52.30GPU estimate0.21
44Mercury Decide decision Closed · EfficientDetails53.451.5–55.2 (50 = always safe)47.9$0.071$71.34List price0.48
45d1-Omni-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 2.2BDetails53.250.9–55.6 (50 = always safe)49.8$0.657$657GPU estimate0.06
46LFM2.5-24B-A2B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 24B MoE · 2.5B activeDetails52.450.3–54.5 (50 = always safe)47.7$38.716$38,716GPU estimate324.8
LFM2.5-24B-A2B off · NVIDIA RTX PRO 6000 Blackwell · chat Open · 24B MoE · 2.5B activeDetails45.743.2–48.1 (50 = always safe)61.5$9.132$9,132GPU estimate0.14
47KnowLine-2B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetails51.148.7–53.5 (50 = always safe)54.5$0.049$49.34GPU estimate0.15
48KnowLine 4B Gen5.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetails49.747.0–52.3 (50 = always safe)53.6$0.054$53.72GPU estimate0.28
49Perplexity Decider Flash v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 4.0BDetails49.146.5–51.6 (50 = always safe)59.6$12.267$12,267GPU estimate0.56
50Open-Jev-27B-v1.1 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails48.546.1–50.9 (50 = always safe)65.6$7.185$7,185GPU estimate1.57
51GLiDE 2B NVIDIA DGX Spark (GB10) · decision · late entry Open · 1.9BDetails46.743.6–49.6 (50 = always safe)67.4$0.058$58.31GPU estimate0.33
52KnowLine-0.8B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails45.242.5–47.7 (50 = always safe)71.9$0.049$48.93GPU estimate0.11
53StartLux-Decision-v1.1-0.8B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails44.642.3–46.9 (50 = always safe)64.7$0.054$53.72GPU estimate0.13
54OneJev1.6-0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.9BDetails42.539.3–45.8 (50 = always safe)82.5$0.052$51.90GPU estimate0.10
55GLiDE 0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails41.738.9–44.3 (50 = always safe)80.4$0.051$51.47GPU estimate0.22
56GLiDE 0.4B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetails28.925.6–32.1 (50 = always safe)113.2$0.057$57.11GPU estimate0.13
57GLiClass v4 Edge NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetails21.518.6–24.6 (50 = always safe)133.5$0.174$174GPU estimate1.33
GNUbg weak programDetails67.163.8–70.3 (50 = always safe)21.4–0.01
GNUbg weak programDetails62.359.5–64.8 (50 = always safe)27.0–0.01
GNUbg weak programDetails62.158.2–65.9 (50 = always safe)28.1–0.01
Safest programDetails50.050.0–50.0 (50 = always safe)51.6–0.00
Safest (computed) programDetails50.050.0–50.0 (50 = always safe)––Not measured
Safest (computed) programDetails50.050.0–50.0 (50 = always safe)––Not measured
Random programDetails24.922.3–27.4 (50 = always safe)121.1–0.00
Random (computed) programDetails27.026.7–27.2 (50 = always safe)––Not measured
Random (computed) programDetails27.627.1–28.0 (50 = always safe)––Not measured

No model of this kind on the leaderboard.

Risk judgement: 100 = every call right, 50 = always safe; higher is better. One row per model at its highest reasoning setting, never its best score; late entries included. All setups: every setup of every model. How rows are chosen

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

Rank counts 1, 2, 3 down the rows shown, by the measure shown (with a filter on, among those rows only); it has no tiers: a tier needs a paired test (McNemar, Holm) over positions every model answered, and each league played its own positions, so models in different leagues never answered the same ones. The tier in each row's details and Same positions (preview) come from 200 positions every model there answered. Rank comes from league play; tier comes from the same 200 positions for every model. Ranks show observed scores; small gaps need not be meaningful.

$ per 1,000 decisions: At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid. Each cost's tag opens how it was counted; each row's details name its price basis.

How rows are chosen

Class: by the maker's total parameters, or by list price when the size is undisclosed *

Size by total parameters, as disclosed by the maker (for MoE models the label also shows active parameters): Tiny under 6B · Small 6B–15B · Medium over 15B up to 40B · Large over 40B. Models of undisclosed size, by standard list price per 1M input tokens: Efficient up to $0.50 · Workhorse up to $3 · Flagship over $3.

Learning Ordered: Yes, then No, then not finished; each by the change in risks found. Not a ranking.

The setup tested: the model and its setting. Tap or point at a name for its details.Yes: the change in risks found is above the threshold and right safety falls by no more than 5 pp. No: either fails. Shown once every run is complete; before that, the model's state (running, queued, waiting and why, not tested and the rule).Risks found after the series minus before, in percentage points (pp): points of the share, not a relative percent.Took the risk when it paid, as a share of the positions where it did: before = the mean of its 3 empty runs (no lessons), after = after the series of lessons.Right safety (played safe when safe was right) after minus before, in percentage points. A fall of more than 5 pp makes the result No.What the change must beat: the larger of the model's own spread over its 3 empty runs (max − min of risks found) and 5 pp.
Improved: Yes
–GLM-5.3 low · chat Open · 753B MoEDetailsYes+17.621.3 → 39.0−4.09.0
–Claude Fable 5.1 high · chat Closed · FlagshipDetailsYes+9.454.0 → 63.4−2.15.0
Claude Fable 5.1 low · chat Closed · FlagshipDetailsnot in the learning plan
–Gemini 4 Argon high · chat Closed · FlagshipDetailsYes+8.5Run 253.3 → 61.8−1.55.0
Gemini 4 Argon low · chat Closed · FlagshipDetailsnot in the learning plan
–Gemma 4 31B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetailsYes+8.534.3 → 42.8−1.75.0
Gemma 4 31B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetailsnot in the learning plan
–DeepSeek-V4.1-Flash low · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetailsYes+8.2Run 244.6 → 52.9−3.65.0
DeepSeek-V4.1-Flash low · API · 1017 · chat Open · 552B MoE · 16B activeDetailsnot in the learning plan
DeepSeek-V4.1-Flash off · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetailsnot in the learning plan
DeepSeek-V4.1-Flash off · API · 1017 · chat Open · 552B MoE · 16B activeDetailsnot in the learning plan
–Gemini 3.8 Flash high · chat Closed · WorkhorseDetailsYes+8.238.0 → 46.2−1.95.0
Gemini 3.8 Flash low · chat Closed · WorkhorseDetailsnot in the learning plan
–Claude Sonnet 5.5 high · chat Closed · WorkhorseDetailsYes+7.043.0 → 50.0−2.95.0
Claude Sonnet 5.5 low · chat Closed · WorkhorseDetailsnot in the learning plan
–Claude Opus 5.5 high · chat Closed · FlagshipDetailsYes+6.747.3 → 54.0+1.15.0
Claude Opus 5.5 low · chat Closed · FlagshipDetailsnot in the learning plan
Improved: No
–Qwen3.8-27B low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetailsNoright safety fell past the limit+19.032.0 → 51.0−7.1past the −5 limit5.0
Qwen3.8-27B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetailsnot in the learning plan
–Kimi K3.1 Pro off · chat Open · 2.8T MoE · 104B activeDetailsNo+4.841.0 → 45.8+1.25.0
–Qwen3.8 Flash off · chat Closed · EfficientDetailsNo+2.733.3 → 36.0+1.55.0
Not finished
–GPT-6 Astra high · chat · late entry Closed · FlagshipDetailspendingN/AN/A → N/AN/AN/A
GPT-6 Astra low · chat Closed · FlagshipDetailsnot in the learning plan
GPT-6 Astra medium · chat Closed · FlagshipDetailsnot in the learning plan
–GPT-6.1 Sol high · chat Closed · WorkhorseDetailspending · numbers interim+5.259.3 → 64.5+2.78.0
GPT-6.1 Sol low · chat Closed · WorkhorseDetailsnot in the learning plan
–Grok 4.7 high · chat Closed · WorkhorseDetailspending · numbers interim+8.248.5 → 56.7−5.2past the −5 limitN/A
Grok 4.7 low · chat Closed · WorkhorseDetailsnot in the learning plan
–d1-ultra decision Closed · EfficientDetailsnot tested (conservative context guard)
–Mammoth-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetailswaiting (Ümit's D5 decision)
–REAX 1.3 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 57BDetailswaiting (Ümit's D5 decision)
–Jev 1.14 decision Closed · EfficientDetailsnot tested (conservative context guard)
–Drex Max v1.1 decision · late entry Closed · Efficient (provisional)Detailsnotes: complete · worst: waiting (cash approval)
–GLiDE v1.2 · NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetailsnotes: stopped (its budget cap was reached) · worst: complete
–Perplexity Decider 27B v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetailscomplete
–Elephant-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 9.0BDetailsnotes: waiting (Ümit's D5 decision) · worst: complete
–Cygnet NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 12BDetailswaiting (its layer never sees the lessons)
–SubQ Snap 1.1 decision · late entry Closed · EfficientDetailscomplete
–d1-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 1.9BDetailsnot started (no backgammon route yet; no league record)
–Winnow-12B NVIDIA RTX PRO 6000 Blackwell · decision Open · 12BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–StartLux-Decision-v1.1-2B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–LFM2.5-24B-A2B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 24B MoE · 2.5B activeDetailspending · numbers interim−8.429.7 → 21.2+0.311.0
LFM2.5-24B-A2B off · NVIDIA RTX PRO 6000 Blackwell · chat Open · 24B MoE · 2.5B activeDetailsnot in the learning plan
–KnowLine-2B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–KnowLine 4B Gen5.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–Perplexity Decider Flash v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 4.0BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–GLiDE 2B NVIDIA DGX Spark (GB10) · decision · late entry Open · 1.9BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–KnowLine-0.8B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–StartLux-Decision-v1.1-0.8B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–OneJev1.6-0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.9BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–GLiDE 0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–GLiDE 0.4B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
–GLiClass v4 Edge NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetailsnotes: waiting (cash approval) · worst: waiting (its render check)
Not in the learning plan
–Muse Ember max · chat · late entry Closed · EfficientDetailsnot in the learning plan
–Gemini Decision 1 20261111 · decision · late entry Closed · EfficientDetailsnot in the learning plan
–Drex Max NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetailsnot in the learning plan
–Sieve 31B v3 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetailsnot in the learning plan
–KnowLine-31B-Gen1.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetailsnot in the learning plan
–GLiDE API · decision Closed · EfficientDetailsnot in the learning plan
–Turion Jev 31B NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetailsnot in the learning plan
–Jev 1.13 decision Closed · EfficientDetailsnot in the learning plan
–Drex 2.0 API · decision · late entry Open · 35B MoE · 3.0B activeDetailsnot in the learning plan
Drex 2.0 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 35B MoE · 3.0B activeDetailsnot in the learning plan
–celeris-1-decision decision Closed · EfficientDetailsnot in the learning plan
–Clef NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetailsnot in the learning plan
–StartLux-Decision-v1.1-4B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetailsnot in the learning plan
–Perplexity Decider 27B v1.1 NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetailsnot in the learning plan
–SubQ Snap Preview decision Closed · EfficientDetailsnot in the learning plan
–Cerno 9B v1.2 NVIDIA DGX Spark (GB10) · decision · late entry Open · 9.0BDetailsnot in the learning plan
–GPT-6 Luna Decisions API · decision Closed · EfficientDetailsnot in the learning plan
–Mercury Decide decision Closed · EfficientDetailsnot in the learning plan
–d1-Omni-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 2.2BDetailsnot in the learning plan
–Open-Jev-27B-v1.1 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetailsnot in the learning plan
GNUbg weak programDetailsprogram · not in the learning plan
GNUbg weak programDetailsprogram · not in the learning plan
GNUbg weak programDetailsprogram · not in the learning plan
Safest programDetailsprogram · not in the learning plan
Safest (computed) programDetailsprogram · not in the learning plan
Safest (computed) programDetailsprogram · not in the learning plan
Random programDetailsprogram · not in the learning plan
Random (computed) programDetailsprogram · not in the learning plan
Random (computed) programDetailsprogram · not in the learning plan

No model of this kind on the leaderboard.

Every model on the leaderboard: its verdict after lessons where one is published (Run 2 where it replaced the original), else its state in the learning lanes, as their status files give it. Did they improve?

Result: Yes or No once every run is complete; the change, right safety change and threshold are in points (pp). *

Threshold: the larger of the model's own spread over its three empty runs (max − min) and 5 points; right safety may fall by at most 5 points.

Class: by the maker's total parameters, or by list price when the size is undisclosed *

Size by total parameters, as disclosed by the maker (for MoE models the label also shows active parameters): Tiny under 6B · Small 6B–15B · Medium over 15B up to 40B · Large over 40B. Models of undisclosed size, by standard list price per 1M input tokens: Efficient up to $0.50 · Workhorse up to $3 · Flagship over $3.

Error rate (PR) all leagues · 95% interval · lower is better

View
  1. 1GPT-6 Astra high · chat · late entry8.8
  2. GPT-6 Astra medium · chat11.1
  3. GPT-6 Astra low · chat11.8
  4. 2GPT-6.1 Sol high · chat9.9
  5. GPT-6.1 Sol low · chat12.1
  6. 3Gemini 4 Argon high · chat12.6
  7. Gemini 4 Argon low · chat16.5
  8. 4Claude Fable 5.1 high · chat14.0
  9. Claude Fable 5.1 low · chat17.9
  10. 5Claude Opus 5.5 high · chat14.4
  11. Claude Opus 5.5 low · chat15.4
  12. 6Gemini 3.8 Flash high · chat14.9
  13. Gemini 3.8 Flash low · chat16.3
  14. 7Claude Sonnet 5.5 high · chat16.7
  15. Claude Sonnet 5.5 low · chat20.1
  16. 8Grok 4.7 high · chat18.8
  17. Grok 4.7 low · chat19.1
  18. 9DeepSeek-V4.1-Flash low · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat19.0
  19. DeepSeek-V4.1-Flash low · API · 1017 · chat17.1
  20. DeepSeek-V4.1-Flash off · API · 1017 · chat35.2
  21. DeepSeek-V4.1-Flash off · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat35.6
  22. 10Gemma 4 31B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat19.6
  23. Gemma 4 31B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat22.5
  24. 11Muse Ember max · chat · late entry20.1
  25. 12KnowLine-31B-Gen1.5 NVIDIA DGX Spark (GB10) · decision · late entry26.2
  26. 13Sieve 31B v3 NVIDIA DGX Spark (GB10) · decision · late entry27.5
  27. 14REAX 1.3 NVIDIA RTX PRO 6000 Blackwell · decision · late entry27.8
  28. 15Gemini Decision 1 20261111 · decision · late entry28.0
  29. 16Kimi K3.1 Pro off · chat28.0
  30. 17Mammoth-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry28.5
  31. 18Qwen3.8-27B low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat29.0
  32. Qwen3.8-27B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat45.0
  33. 19Turion Jev 31B NVIDIA DGX Spark (GB10) · decision · late entry29.7
  34. 20d1-ultra decision30.3
  35. 21Drex Max NVIDIA RTX PRO 6000 Blackwell · decision · late entry30.4
  36. 22GLM-5.3 low · chat30.9
  37. 23Drex Max v1.1 decision · late entry30.9
  38. 24Jev 1.14 decision31.7
  39. 25GLiDE API · decision33.2
  40. 26Qwen3.8 Flash off · chat33.6
  41. 27Jev 1.13 decision34.4
  42. 28GLiDE v1.2 · NVIDIA RTX PRO 6000 Blackwell · decision35.0
  43. 29celeris-1-decision decision35.6
  44. 30Clef NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision36.5
  45. 31Perplexity Decider 27B v1.2 NVIDIA RTX PRO 6000 Blackwell · decision36.9
  46. 32Drex 2.0 API · decision · late entry38.4
  47. Drex 2.0 NVIDIA RTX PRO 6000 Blackwell · decision · late entry42.9
  48. 33Cygnet NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision38.6
  49. 34Perplexity Decider 27B v1.1 NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision38.8
  50. 35StartLux-Decision-v1.1-4B-Instant NVIDIA DGX Spark (GB10) · decision · late entry40.1
  51. 36SubQ Snap 1.1 decision · late entry40.2
  52. 37GPT-6 Luna Decisions API · decision43.3
  53. 38SubQ Snap Preview decision43.9
  54. 39Elephant-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry44.6
  55. 40d1-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry47.1
  56. 41LFM2.5-24B-A2B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat47.7
  57. LFM2.5-24B-A2B off · NVIDIA RTX PRO 6000 Blackwell · chat61.5
  58. 42Mercury Decide decision47.9
  59. 43Cerno 9B v1.2 NVIDIA DGX Spark (GB10) · decision · late entry48.5
  60. 44Winnow-12B NVIDIA RTX PRO 6000 Blackwell · decision49.7
  61. 45d1-Omni-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry49.8
  62. 46StartLux-Decision-v1.1-2B-Instant NVIDIA DGX Spark (GB10) · decision · late entry50.0
  63. 47KnowLine 4B Gen5.5 NVIDIA DGX Spark (GB10) · decision · late entry53.6
  64. 48KnowLine-2B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry54.5
  65. 49Perplexity Decider Flash v1.2 NVIDIA RTX PRO 6000 Blackwell · decision59.6
  66. 50StartLux-Decision-v1.1-0.8B-Instant NVIDIA DGX Spark (GB10) · decision · late entry64.7
  67. 51Open-Jev-27B-v1.1 NVIDIA RTX PRO 6000 Blackwell · decision65.6
  68. 52GLiDE 2B NVIDIA DGX Spark (GB10) · decision · late entry67.4
  69. 53KnowLine-0.8B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry71.9
  70. 54GLiDE 0.8B NVIDIA DGX Spark (GB10) · decision · late entry80.4
  71. 55OneJev1.6-0.8B NVIDIA DGX Spark (GB10) · decision · late entry82.5
  72. 56GLiDE 0.4B NVIDIA DGX Spark (GB10) · decision · late entry113.2
  73. 57GLiClass v4 Edge NVIDIA DGX Spark (GB10) · decision · late entry133.5
  74. GNUbg weak program21.4
  75. GNUbg weak program27.0
  76. GNUbg weak program28.1
  77. Safest program51.6
  78. Safest (computed) programNo value
  79. Safest (computed) programNo value
  80. Random program121.1
  81. Random (computed) programNo value
  82. Random (computed) programNo value

No model of this kind on the leaderboard.

  • OpenAI
  • Google
  • Anthropic
  • xAI
  • DeepSeek
  • Meta
  • PelaAI
  • Sthānika AI
  • REAX
  • Moonshot
  • Base Labs
  • Alibaba
  • Turion.AI
  • Liquid AI
  • Nace.AI
  • Z.ai
  • TypeSafe
  • Fastino
  • Celeris
  • Cloudflare
  • Perplexity
  • Blockbrain Labs
  • StartLux
  • Subquadratic
  • Inception
  • Apertix AI
  • EldanRing
  • Zefan Cai
  • OmniJev
  • Knowledgator
  • program
Counts 1, 2, 3 down the rows shown, by the measure chosen; no tiers.The model and its setting. Tap or point at a name for its details.500 × mean 2-ply equity loss per decision, the checker-play performance rating (PR) backgammon players use; lower is better.100 = every call right, 50 = always safe; higher is better.Dollars per 1,000 decisions at each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid. Lower is better.Median request seconds per decision (waits between retries not included). Lower is better.
1GPT-6 Astra high · chat · late entry Closed · FlagshipDetails8.87.4–10.374.4$59.356$59,356List price9.83
GPT-6 Astra medium · chat Closed · FlagshipDetails11.19.5–12.974.7$50.268$50,268List price4.40
GPT-6 Astra low · chat Closed · FlagshipDetails11.89.5–14.275.8$49.737$49,737List price3.94
2GPT-6.1 Sol high · chat Closed · WorkhorseDetails9.98.6–11.372.2$15.573$15,573List price13.1
GPT-6.1 Sol low · chat Closed · WorkhorseDetails12.110.5–13.772.7$10.228$10,228List price4.33
3Gemini 4 Argon high · chat Closed · FlagshipDetails12.610.5–14.874.1$64.401$64,401List price27.5
Gemini 4 Argon low · chat Closed · FlagshipDetails16.513.7–19.569.2$10.204$10,204List price3.82
4Claude Fable 5.1 high · chat Closed · FlagshipDetails14.011.5–16.372.3$136.268$136,268List price12.7
Claude Fable 5.1 low · chat Closed · FlagshipDetails17.915.9–20.167.8$62.613$62,613List price1.54
5Claude Opus 5.5 high · chat Closed · FlagshipDetails14.412.1–17.070.0$36.400$36,400List price6.35
Claude Opus 5.5 low · chat Closed · FlagshipDetails15.413.1–18.066.9$29.062$29,062List price3.92
6Gemini 3.8 Flash high · chat Closed · WorkhorseDetails14.912.9–17.066.9$22.352$22,352List price9.59
Gemini 3.8 Flash low · chat Closed · WorkhorseDetails16.314.5–18.265.1$4.123$4,123List price2.16
7Claude Sonnet 5.5 high · chat Closed · WorkhorseDetails16.714.9–18.768.2$19.847$19,847List price6.63
Claude Sonnet 5.5 low · chat Closed · WorkhorseDetails20.117.8–22.463.5$13.541$13,541List price1.51
8Grok 4.7 high · chat Closed · WorkhorseDetails18.815.9–21.868.5$117.379$117,379List price122.3
Grok 4.7 low · chat Closed · WorkhorseDetails19.116.9–21.467.7$32.813$32,813List price24.3
9DeepSeek-V4.1-Flash low · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetails19.016.6–21.369.0$47.092$47,092GPU estimate43.8
DeepSeek-V4.1-Flash low · API · 1017 · chat Open · 552B MoE · 16B activeDetails17.115.1–19.268.5$4.869$4,869List price22.3
DeepSeek-V4.1-Flash off · API · 1017 · chat Open · 552B MoE · 16B activeDetails35.231.8–38.858.5$0.244$244List price0.84
DeepSeek-V4.1-Flash off · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetails35.631.7–39.759.3$0.728$728GPU estimate0.48
10Gemma 4 31B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetails19.617.5–21.764.5$34.943$34,943GPU estimate120.6
Gemma 4 31B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetails22.519.9–25.362.6$8.215$8,215GPU estimate0.31
11Muse Ember max · chat · late entry Closed · EfficientDetails20.117.6–22.766.0$6.863$6,863Billed15.3
12KnowLine-31B-Gen1.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails26.223.0–29.461.6–3.45
13Sieve 31B v3 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails27.524.1–31.662.7$0.794$794GPU estimate11.1
14REAX 1.3 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 57BDetails27.824.9–30.660.5$1.545$1,545GPU estimate1.74
15Gemini Decision 1 20261111 · decision · late entry Closed · EfficientDetails28.025.0–31.164.7$0.382$382Billed0.24
16Kimi K3.1 Pro off · chat Open · 2.8T MoE · 104B activeDetails28.024.5–31.865.9$3.982$3,982List price2.11
17Mammoth-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetails28.526.2–31.163.2$0.999$999GPU estimate0.72
18Qwen3.8-27B low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetails29.026.5–31.862.0$39.023$39,023GPU estimate171.1
Qwen3.8-27B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetails45.040.7–49.957.6$4.551$4,551GPU estimate1.48
19Turion Jev 31B NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails29.726.5–33.359.6$0.293$293GPU estimate3.95
20d1-ultra decision Closed · EfficientDetails30.327.3–33.464.0$0.145$145List price0.23
21Drex Max NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetails30.427.4–33.462.9$1.285$1,285GPU estimate0.68
22GLM-5.3 low · chat Open · 753B MoEDetails30.927.8–34.061.7$2.917$2,917List price1.75
23Drex Max v1.1 decision · late entry Closed · Efficient (provisional)Details30.928.2–33.759.1$0.265$265List price0.44
24Jev 1.14 decision Closed · EfficientDetails31.728.3–35.759.3$0.173$173List price0.19
25GLiDE API · decision Closed · EfficientDetails33.230.2–36.460.4$0.153$153List price0.48
26Qwen3.8 Flash off · chat Closed · EfficientDetails33.630.8–36.759.5$0.284$284List price1.70
27Jev 1.13 decision Closed · EfficientDetails34.431.0–38.059.5$0.166$166List price0.22
28GLiDE v1.2 · NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails35.031.6–39.058.8$10.301$10,301GPU estimate1.41
29celeris-1-decision decision Closed · EfficientDetails35.632.0–39.457.8$0.150$150List price0.17
30Clef NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetails36.533.5–39.657.5$5.736$5,736GPU estimate1.13
31Perplexity Decider 27B v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails36.933.9–40.157.9$12.227$12,227GPU estimate0.94
32Drex 2.0 API · decision · late entry Open · 35B MoE · 3.0B activeDetails38.435.6–41.458.1$0.258$258Billed0.42
Drex 2.0 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 35B MoE · 3.0B activeDetails42.939.1–47.153.0$0.782$782GPU estimate0.40
33Cygnet NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 12BDetails38.635.4–41.957.1$11.168$11,168GPU estimate0.37
34Perplexity Decider 27B v1.1 NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetails38.835.3–42.556.9$6.593$6,593GPU estimate1.97
35StartLux-Decision-v1.1-4B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetails40.136.7–43.757.2$0.072$71.58GPU estimate0.44
36SubQ Snap 1.1 decision · late entry Closed · EfficientDetails40.236.7–43.856.7$0.150$150Billed0.23
37GPT-6 Luna Decisions API · decision Closed · EfficientDetails43.339.7–46.955.7$0.335$335List price0.23
38SubQ Snap Preview decision Closed · EfficientDetails43.939.3–48.856.5$0.193$193List price0.24
39Elephant-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 9.0BDetails44.640.3–48.957.7$0.685$685GPU estimate0.26
40d1-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 1.9BDetails47.142.8–51.555.2$0.467$467GPU estimate0.06
41LFM2.5-24B-A2B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 24B MoE · 2.5B activeDetails47.744.1–51.252.4$38.716$38,716GPU estimate324.8
LFM2.5-24B-A2B off · NVIDIA RTX PRO 6000 Blackwell · chat Open · 24B MoE · 2.5B activeDetails61.556.3–66.845.7$9.132$9,132GPU estimate0.14
42Mercury Decide decision Closed · EfficientDetails47.943.3–52.953.4$0.071$71.34List price0.48
43Cerno 9B v1.2 NVIDIA DGX Spark (GB10) · decision · late entry Open · 9.0BDetails48.544.5–52.656.1$0.139$139GPU estimate1.96
44Winnow-12B NVIDIA RTX PRO 6000 Blackwell · decision Open · 12BDetails49.745.9–53.655.2$9.615$9,615GPU estimate2.46
45d1-Omni-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 2.2BDetails49.843.5–57.453.2$0.657$657GPU estimate0.06
46StartLux-Decision-v1.1-2B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetails50.045.4–54.853.8$0.052$52.30GPU estimate0.21
47KnowLine 4B Gen5.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetails53.648.8–58.649.7$0.054$53.72GPU estimate0.28
48KnowLine-2B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetails54.548.9–60.851.1$0.049$49.34GPU estimate0.15
49Perplexity Decider Flash v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 4.0BDetails59.654.9–64.549.1$12.267$12,267GPU estimate0.56
50StartLux-Decision-v1.1-0.8B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails64.759.3–70.344.6$0.054$53.72GPU estimate0.13
51Open-Jev-27B-v1.1 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails65.659.7–71.948.5$7.185$7,185GPU estimate1.57
52GLiDE 2B NVIDIA DGX Spark (GB10) · decision · late entry Open · 1.9BDetails67.461.0–74.546.7$0.058$58.31GPU estimate0.33
53KnowLine-0.8B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails71.966.2–77.945.2$0.049$48.93GPU estimate0.11
54GLiDE 0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails80.471.8–89.341.7$0.051$51.47GPU estimate0.22
55OneJev1.6-0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.9BDetails82.575.2–90.842.5$0.052$51.90GPU estimate0.10
56GLiDE 0.4B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetails113.2102.1–124.128.9$0.057$57.11GPU estimate0.13
57GLiClass v4 Edge NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetails133.5118.5–148.821.5$0.174$174GPU estimate1.33
GNUbg weak programDetails21.418.8–24.367.1–0.01
GNUbg weak programDetails27.024.1–29.862.3–0.01
GNUbg weak programDetails28.125.0–31.462.1–0.01
Safest programDetails51.647.4–56.250.0–0.00
Safest (computed) programDetails–50.0–Not measured
Safest (computed) programDetails–50.0–Not measured
Random programDetails121.1113.3–128.524.9–0.00
Random (computed) programDetails–27.0–Not measured
Random (computed) programDetails–27.6–Not measured

No model of this kind on the leaderboard.

One row per model at its highest reasoning setting, never its best score; late entries included. All setups: every setup of every model. How rows are chosen

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

Rank counts 1, 2, 3 down the rows shown, by the measure shown (with a filter on, among those rows only); it has no tiers: a tier needs a paired test (McNemar, Holm) over positions every model answered, and each league played its own positions, so models in different leagues never answered the same ones. The tier in each row's details and Same positions (preview) come from 200 positions every model there answered. Rank comes from league play; tier comes from the same 200 positions for every model. Ranks show observed scores; small gaps need not be meaningful.

$ per 1,000 decisions: At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid. Each cost's tag opens how it was counted; each row's details name its price basis.

How rows are chosen

Class: by the maker's total parameters, or by list price when the size is undisclosed *

Size by total parameters, as disclosed by the maker (for MoE models the label also shows active parameters): Tiny under 6B · Small 6B–15B · Medium over 15B up to 40B · Large over 40B. Models of undisclosed size, by standard list price per 1M input tokens: Efficient up to $0.50 · Workhorse up to $3 · Flagship over $3.

Cost per 1,000 decisions all leagues · lower is better

View
  1. 1KnowLine-0.8B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry$0.049
  2. 2KnowLine-2B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry$0.049
  3. 3GLiDE 0.8B NVIDIA DGX Spark (GB10) · decision · late entry$0.051
  4. 4OneJev1.6-0.8B NVIDIA DGX Spark (GB10) · decision · late entry$0.052
  5. 5StartLux-Decision-v1.1-2B-Instant NVIDIA DGX Spark (GB10) · decision · late entry$0.052
  6. 6StartLux-Decision-v1.1-0.8B-Instant NVIDIA DGX Spark (GB10) · decision · late entry$0.054
  7. 7KnowLine 4B Gen5.5 NVIDIA DGX Spark (GB10) · decision · late entry$0.054
  8. 8GLiDE 0.4B NVIDIA DGX Spark (GB10) · decision · late entry$0.057
  9. 9GLiDE 2B NVIDIA DGX Spark (GB10) · decision · late entry$0.058
  10. 10Mercury Decide decision$0.071
  11. 11StartLux-Decision-v1.1-4B-Instant NVIDIA DGX Spark (GB10) · decision · late entry$0.072
  12. 12Cerno 9B v1.2 NVIDIA DGX Spark (GB10) · decision · late entry$0.139
  13. 13d1-ultra decision$0.145
  14. 14SubQ Snap 1.1 decision · late entry$0.150
  15. 15celeris-1-decision decision$0.150
  16. 16GLiDE API · decision$0.153
  17. 17Jev 1.13 decision$0.166
  18. 18Jev 1.14 decision$0.173
  19. 19GLiClass v4 Edge NVIDIA DGX Spark (GB10) · decision · late entry$0.174
  20. 20SubQ Snap Preview decision$0.193
  21. 21Drex 2.0 API · decision · late entry$0.258
  22. Drex 2.0 NVIDIA RTX PRO 6000 Blackwell · decision · late entry$0.782
  23. 22Drex Max v1.1 decision · late entry$0.265
  24. 23Qwen3.8 Flash off · chat$0.284
  25. 24Turion Jev 31B NVIDIA DGX Spark (GB10) · decision · late entry$0.293
  26. 25GPT-6 Luna Decisions API · decision$0.335
  27. 26Gemini Decision 1 20261111 · decision · late entry$0.382
  28. 27d1-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry$0.467
  29. 28d1-Omni-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry$0.657
  30. 29Elephant-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry$0.685
  31. 30Sieve 31B v3 NVIDIA DGX Spark (GB10) · decision · late entry$0.794
  32. 31Mammoth-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry$0.999
  33. 32Drex Max NVIDIA RTX PRO 6000 Blackwell · decision · late entry$1.285
  34. 33REAX 1.3 NVIDIA RTX PRO 6000 Blackwell · decision · late entry$1.545
  35. 34GLM-5.3 low · chat$2.917
  36. 35Kimi K3.1 Pro off · chat$3.982
  37. 36Clef NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision$5.736
  38. 37Perplexity Decider 27B v1.1 NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision$6.593
  39. 38Muse Ember max · chat · late entry$6.863
  40. 39Open-Jev-27B-v1.1 NVIDIA RTX PRO 6000 Blackwell · decision$7.185
  41. 40Winnow-12B NVIDIA RTX PRO 6000 Blackwell · decision$9.615
  42. 41GLiDE v1.2 · NVIDIA RTX PRO 6000 Blackwell · decision$10.301
  43. 42Cygnet NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision$11.168
  44. 43Perplexity Decider 27B v1.2 NVIDIA RTX PRO 6000 Blackwell · decision$12.227
  45. 44Perplexity Decider Flash v1.2 NVIDIA RTX PRO 6000 Blackwell · decision$12.267
  46. 45GPT-6.1 Sol high · chat$15.573
  47. GPT-6.1 Sol low · chat$10.228
  48. 46Claude Sonnet 5.5 high · chat$19.847
  49. Claude Sonnet 5.5 low · chat$13.541
  50. 47Gemini 3.8 Flash high · chat$22.352
  51. Gemini 3.8 Flash low · chat$4.123
  52. 48Gemma 4 31B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat$34.943
  53. Gemma 4 31B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat$8.215
  54. 49Claude Opus 5.5 high · chat$36.400
  55. Claude Opus 5.5 low · chat$29.062
  56. 50LFM2.5-24B-A2B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat$38.716
  57. LFM2.5-24B-A2B off · NVIDIA RTX PRO 6000 Blackwell · chat$9.132
  58. 51Qwen3.8-27B low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat$39.023
  59. Qwen3.8-27B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat$4.551
  60. 52DeepSeek-V4.1-Flash low · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat$47.092
  61. DeepSeek-V4.1-Flash off · API · 1017 · chat$0.244
  62. DeepSeek-V4.1-Flash off · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat$0.728
  63. DeepSeek-V4.1-Flash low · API · 1017 · chat$4.869
  64. 53GPT-6 Astra high · chat · late entry$59.356
  65. GPT-6 Astra low · chat$49.737
  66. GPT-6 Astra medium · chat$50.268
  67. 54Gemini 4 Argon high · chat$64.401
  68. Gemini 4 Argon low · chat$10.204
  69. 55Grok 4.7 high · chat$117.379
  70. Grok 4.7 low · chat$32.813
  71. 56Claude Fable 5.1 high · chat$136.268
  72. Claude Fable 5.1 low · chat$62.613
  73. –KnowLine-31B-Gen1.5 NVIDIA DGX Spark (GB10) · decision · late entryNo value
  74. GNUbg weak programNo value
  75. GNUbg weak programNo value
  76. GNUbg weak programNo value
  77. Safest programNo value
  78. Safest (computed) programNo value
  79. Safest (computed) programNo value
  80. Random programNo value
  81. Random (computed) programNo value
  82. Random (computed) programNo value

No model of this kind on the leaderboard.

  • PelaAI
  • Fastino
  • OmniJev
  • StartLux
  • Inception
  • Apertix AI
  • Liquid AI
  • Subquadratic
  • Celeris
  • TypeSafe
  • Knowledgator
  • Nace.AI
  • Alibaba
  • Turion.AI
  • OpenAI
  • Google
  • Base Labs
  • Sthānika AI
  • REAX
  • Z.ai
  • Moonshot
  • Cloudflare
  • Perplexity
  • Meta
  • Zefan Cai
  • EldanRing
  • Blockbrain Labs
  • Anthropic
  • DeepSeek
  • xAI
  • program
Counts 1, 2, 3 down the rows shown, by the measure chosen; no tiers.The model and its setting. Tap or point at a name for its details.Dollars per 1,000 decisions at each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid. Lower is better.100 = every call right, 50 = always safe; higher is better.500 × mean 2-ply equity loss per decision, the checker-play performance rating (PR) backgammon players use; lower is better.Median request seconds per decision (waits between retries not included). Lower is better.
1KnowLine-0.8B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails$0.049$48.93GPU estimate45.271.90.11
2KnowLine-2B-Gen1 NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetails$0.049$49.34GPU estimate51.154.50.15
3GLiDE 0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails$0.051$51.47GPU estimate41.780.40.22
4OneJev1.6-0.8B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.9BDetails$0.052$51.90GPU estimate42.582.50.10
5StartLux-Decision-v1.1-2B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 2.0BDetails$0.052$52.30GPU estimate53.850.00.21
6StartLux-Decision-v1.1-0.8B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.8BDetails$0.054$53.72GPU estimate44.664.70.13
7KnowLine 4B Gen5.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetails$0.054$53.72GPU estimate49.753.60.28
8GLiDE 0.4B NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetails$0.057$57.11GPU estimate28.9113.20.13
9GLiDE 2B NVIDIA DGX Spark (GB10) · decision · late entry Open · 1.9BDetails$0.058$58.31GPU estimate46.767.40.33
10Mercury Decide decision Closed · EfficientDetails$0.071$71.34List price53.447.90.48
11StartLux-Decision-v1.1-4B-Instant NVIDIA DGX Spark (GB10) · decision · late entry Open · 4.0BDetails$0.072$71.58GPU estimate57.240.10.44
12Cerno 9B v1.2 NVIDIA DGX Spark (GB10) · decision · late entry Open · 9.0BDetails$0.139$139GPU estimate56.148.51.96
13d1-ultra decision Closed · EfficientDetails$0.145$145List price64.030.30.23
14SubQ Snap 1.1 decision · late entry Closed · EfficientDetails$0.150$150Billed56.740.20.23
15celeris-1-decision decision Closed · EfficientDetails$0.150$150List price57.835.60.17
16GLiDE API · decision Closed · EfficientDetails$0.153$153List price60.433.20.48
17Jev 1.13 decision Closed · EfficientDetails$0.166$166List price59.534.40.22
18Jev 1.14 decision Closed · EfficientDetails$0.173$173List price59.331.70.19
19GLiClass v4 Edge NVIDIA DGX Spark (GB10) · decision · late entry Open · 0.4BDetails$0.174$174GPU estimate21.5133.51.33
20SubQ Snap Preview decision Closed · EfficientDetails$0.193$193List price56.543.90.24
21Drex 2.0 API · decision · late entry Open · 35B MoE · 3.0B activeDetails$0.258$258Billed58.138.40.42
Drex 2.0 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 35B MoE · 3.0B activeDetails$0.782$782GPU estimate53.042.90.40
22Drex Max v1.1 decision · late entry Closed · Efficient (provisional)Details$0.265$265List price59.130.90.44
23Qwen3.8 Flash off · chat Closed · EfficientDetails$0.284$284List price59.533.61.70
24Turion Jev 31B NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails$0.293$293GPU estimate59.629.73.95
25GPT-6 Luna Decisions API · decision Closed · EfficientDetails$0.335$335List price55.743.30.23
26Gemini Decision 1 20261111 · decision · late entry Closed · EfficientDetails$0.382$382Billed64.728.00.24
27d1-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 1.9BDetails$0.467$467GPU estimate55.247.10.06
28d1-Omni-2B NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 2.2BDetails$0.657$657GPU estimate53.249.80.06
29Elephant-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 9.0BDetails$0.685$685GPU estimate57.744.60.26
30Sieve 31B v3 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails$0.794$794GPU estimate62.727.511.1
31Mammoth-1 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetails$0.999$999GPU estimate63.228.50.72
32Drex Max NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 27BDetails$1.285$1,285GPU estimate62.930.40.68
33REAX 1.3 NVIDIA RTX PRO 6000 Blackwell · decision · late entry Open · 57BDetails$1.545$1,545GPU estimate60.527.81.74
34GLM-5.3 low · chat Open · 753B MoEDetails$2.917$2,917List price61.730.91.75
35Kimi K3.1 Pro off · chat Open · 2.8T MoE · 104B activeDetails$3.982$3,982List price65.928.02.11
36Clef NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetails$5.736$5,736GPU estimate57.536.51.13
37Perplexity Decider 27B v1.1 NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 27BDetails$6.593$6,593GPU estimate56.938.81.97
38Muse Ember max · chat · late entry Closed · EfficientDetails$6.863$6,863Billed66.020.115.3
39Open-Jev-27B-v1.1 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails$7.185$7,185GPU estimate48.565.61.57
40Winnow-12B NVIDIA RTX PRO 6000 Blackwell · decision Open · 12BDetails$9.615$9,615GPU estimate55.249.72.46
41GLiDE v1.2 · NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails$10.301$10,301GPU estimate58.835.01.41
42Cygnet NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · decision Open · 12BDetails$11.168$11,168GPU estimate57.138.60.37
43Perplexity Decider 27B v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 27BDetails$12.227$12,227GPU estimate57.936.90.94
44Perplexity Decider Flash v1.2 NVIDIA RTX PRO 6000 Blackwell · decision Open · 4.0BDetails$12.267$12,267GPU estimate49.159.60.56
45GPT-6.1 Sol high · chat Closed · WorkhorseDetails$15.573$15,573List price72.29.913.1
GPT-6.1 Sol low · chat Closed · WorkhorseDetails$10.228$10,228List price72.712.14.33
46Claude Sonnet 5.5 high · chat Closed · WorkhorseDetails$19.847$19,847List price68.216.76.63
Claude Sonnet 5.5 low · chat Closed · WorkhorseDetails$13.541$13,541List price63.520.11.51
47Gemini 3.8 Flash high · chat Closed · WorkhorseDetails$22.352$22,352List price66.914.99.59
Gemini 3.8 Flash low · chat Closed · WorkhorseDetails$4.123$4,123List price65.116.32.16
48Gemma 4 31B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetails$34.943$34,943GPU estimate64.519.6120.6
Gemma 4 31B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 31BDetails$8.215$8,215GPU estimate62.622.50.31
49Claude Opus 5.5 high · chat Closed · FlagshipDetails$36.400$36,400List price70.014.46.35
Claude Opus 5.5 low · chat Closed · FlagshipDetails$29.062$29,062List price66.915.43.92
50LFM2.5-24B-A2B on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 24B MoE · 2.5B activeDetails$38.716$38,716GPU estimate52.447.7324.8
LFM2.5-24B-A2B off · NVIDIA RTX PRO 6000 Blackwell · chat Open · 24B MoE · 2.5B activeDetails$9.132$9,132GPU estimate45.761.50.14
51Qwen3.8-27B low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetails$39.023$39,023GPU estimate62.029.0171.1
Qwen3.8-27B off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · chat Open · 27BDetails$4.551$4,551GPU estimate57.645.01.48
52DeepSeek-V4.1-Flash low · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetails$47.092$47,092GPU estimate69.019.043.8
DeepSeek-V4.1-Flash off · API · 1017 · chat Open · 552B MoE · 16B activeDetails$0.244$244List price58.535.20.84
DeepSeek-V4.1-Flash off · API + NVIDIA RTX PRO 6000 Blackwell · first release · chat Open · 552B MoE · 16B activeDetails$0.728$728GPU estimate59.335.60.48
DeepSeek-V4.1-Flash low · API · 1017 · chat Open · 552B MoE · 16B activeDetails$4.869$4,869List price68.517.122.3
53GPT-6 Astra high · chat · late entry Closed · FlagshipDetails$59.356$59,356List price74.48.89.83
GPT-6 Astra low · chat Closed · FlagshipDetails$49.737$49,737List price75.811.83.94
GPT-6 Astra medium · chat Closed · FlagshipDetails$50.268$50,268List price74.711.14.40
54Gemini 4 Argon high · chat Closed · FlagshipDetails$64.401$64,401List price74.112.627.5
Gemini 4 Argon low · chat Closed · FlagshipDetails$10.204$10,204List price69.216.53.82
55Grok 4.7 high · chat Closed · WorkhorseDetails$117.379$117,379List price68.518.8122.3
Grok 4.7 low · chat Closed · WorkhorseDetails$32.813$32,813List price67.719.124.3
56Claude Fable 5.1 high · chat Closed · FlagshipDetails$136.268$136,268List price72.314.012.7
Claude Fable 5.1 low · chat Closed · FlagshipDetails$62.613$62,613List price67.817.91.54
–KnowLine-31B-Gen1.5 NVIDIA DGX Spark (GB10) · decision · late entry Open · 31BDetails–61.626.23.45
GNUbg weak programDetails–67.121.40.01
GNUbg weak programDetails–62.327.00.01
GNUbg weak programDetails–62.128.10.01
Safest programDetails–50.051.60.00
Safest (computed) programDetails–50.0–Not measured
Safest (computed) programDetails–50.0–Not measured
Random programDetails–24.9121.10.00
Random (computed) programDetails–27.0–Not measured
Random (computed) programDetails–27.6–Not measured

No model of this kind on the leaderboard.

One row per model at its highest reasoning setting, never its best score; late entries included. All setups: every setup of every model. How rows are chosen

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

Rank counts 1, 2, 3 down the rows shown, by the measure shown (with a filter on, among those rows only); it has no tiers: a tier needs a paired test (McNemar, Holm) over positions every model answered, and each league played its own positions, so models in different leagues never answered the same ones. The tier in each row's details and Same positions (preview) come from 200 positions every model there answered. Rank comes from league play; tier comes from the same 200 positions for every model. Ranks show observed scores; small gaps need not be meaningful.

$ per 1,000 decisions: At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid. Each cost's tag opens how it was counted; each row's details name its price basis.

How rows are chosen

Class: by the maker's total parameters, or by list price when the size is undisclosed *

Size by total parameters, as disclosed by the maker (for MoE models the label also shows active parameters): Tiny under 6B · Small 6B–15B · Medium over 15B up to 40B · Large over 40B. Models of undisclosed size, by standard list price per 1M input tokens: Efficient up to $0.50 · Workhorse up to $3 · Flagship over $3.

Details of each setup

GPT-6 Astra · OpenAI · chat model · high

Risk judgement 74.4 (95%: 71.8–77.0)

Risks found 53.3% · safe right 95.5%

Took 248 of 465 risks that paid · played safe in 614 of 643 positions where safe paid

Error rate 8.8 (95%: 7.4–10.3) · record 10–1–1 · late entry

Played the original League 3 field after the league finished (late entry)

$59.356 per 1,000 decisions$59.356 per 1M decisions (API-equivalent (subscription)) · 1,489 decisions

Parameters: not disclosed by maker

Closed · Flagship ($10 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Its late-entry page →

Did they improve? pending

GPT-6 Astra · OpenAI · chat model · low

Risk judgement 75.8 (95%: 73.1–78.5)

Risks found 57.2% · safe right 94.4%

Took 226 of 395 risks that paid · played safe in 486 of 515 positions where safe paid

Error rate 11.8 (95%: 9.5–14.2) · record 4–1–6

League 3

$49.737 per 1,000 decisions$49.737 per 1M decisions (API-equivalent (subscription)) · 1,251 decisions

Parameters: not disclosed by maker

Closed · Flagship ($10 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

GPT-6 Astra · OpenAI · chat model · medium

Risk judgement 74.7 (95%: 71.9–77.3)

Risks found 54.9% · safe right 94.4%

Took 245 of 446 risks that paid · played safe in 542 of 574 positions where safe paid

Error rate 11.1 (95%: 9.5–12.9) · record 8–0–3

League 3

$50.268 per 1,000 decisions$50.268 per 1M decisions (API-equivalent (subscription)) · 1,362 decisions

Parameters: not disclosed by maker

Closed · Flagship ($10 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

Gemini 4 Argon · Google · chat model · high

Risk judgement 74.1 (95%: 72.0–76.1)

Risks found 50.8% · safe right 97.4%

Took 224 of 441 risks that paid · played safe in 484 of 497 positions where safe paid

Error rate 12.6 (95%: 10.5–14.8) · record 4–1–6

League 3

$64.401 per 1,000 decisions$64.401 per 1M decisions (API-equivalent (subscription)) · 1,368 decisions

Parameters: not disclosed by maker

Closed · Flagship ($4 per 1M input tokens, standard list price).

Tier (same positions) 1

Released Oct 2026

Model page →

Did they improve? Yes · +8.5 pp · threshold 5.0 pp

Gemini 4 Argon · Google · chat model · low

Risk judgement 69.2 (95%: 66.2–72.3)

Risks found 47.8% · safe right 90.7%

Took 209 of 437 risks that paid · played safe in 543 of 599 positions where safe paid

Error rate 16.5 (95%: 13.7–19.5) · record 3–1–7

League 3

$10.204 per 1,000 decisions$10.204 per 1M decisions (API-equivalent (subscription)) · 1,484 decisions

Parameters: not disclosed by maker

Closed · Flagship ($4 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? not in the learning plan

Claude Fable 5.1 · Anthropic · chat model · high

Risk judgement 72.3 (95%: 69.8–74.7)

Risks found 50.7% · safe right 94.0%

Took 215 of 424 risks that paid · played safe in 529 of 563 positions where safe paid

Error rate 14.0 (95%: 11.5–16.3) · record 6–0–5

League 3

$136.268 per 1,000 decisions$136.268 per 1M decisions (API-equivalent (subscription)) · 1,286 decisions

Parameters: not disclosed by maker

Closed · Flagship ($10 per 1M input tokens, standard list price).

Tier (same positions) 1

Released Sep 2026

Model page →

Did they improve? Yes · +9.4 pp · threshold 5.0 pp

Claude Fable 5.1 · Anthropic · chat model · low

Risk judgement 67.8 (95%: 65.6–69.9)

Risks found 41.0% · safe right 94.6%

Took 270 of 658 risks that paid · played safe in 781 of 826 positions where safe paid

Error rate 17.9 (95%: 15.9–20.1) · record 10–0–8

League 2

$62.613 per 1,000 decisions$62.613 per 1M decisions (API-equivalent (subscription)) · 2,008 decisions

Parameters: not disclosed by maker

Closed · Flagship ($10 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

GPT-6.1 Sol · OpenAI · chat model · high

Risk judgement 72.2 (95%: 69.5–74.9)

Risks found 50.6% · safe right 93.7%

Took 212 of 419 risks that paid · played safe in 568 of 606 positions where safe paid

Error rate 9.9 (95%: 8.6–11.3) · record 7–0–4

League 3

$15.573 per 1,000 decisions$15.573 per 1M decisions (API-equivalent (subscription)) · 1,389 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($2 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? pending

GPT-6.1 Sol · OpenAI · chat model · low

Risk judgement 72.7 (95%: 69.9–75.5)

Risks found 51.7% · safe right 93.8%

Took 198 of 383 risks that paid · played safe in 602 of 642 positions where safe paid

Error rate 12.1 (95%: 10.5–13.7) · record 9–0–2

League 3

$10.228 per 1,000 decisions$10.228 per 1M decisions (API-equivalent (subscription)) · 1,344 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($2 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

Claude Opus 5.5 · Anthropic · chat model · high

Risk judgement 70.0 (95%: 67.3–72.6)

Risks found 43.4% · safe right 96.6%

Took 199 of 459 risks that paid · played safe in 590 of 611 positions where safe paid

Error rate 14.4 (95%: 12.1–17.0) · record 6–0–5

League 3

$36.400 per 1,000 decisions$36.400 per 1M decisions (API-equivalent (subscription)) · 1,477 decisions

Parameters: not disclosed by maker

Closed · Flagship ($4 per 1M input tokens, standard list price).

Tier (same positions) 1

Released Sep 2026

Model page →

Did they improve? Yes · +6.7 pp · threshold 5.0 pp

Claude Opus 5.5 · Anthropic · chat model · low

Risk judgement 66.9 (95%: 64.7–69.2)

Risks found 40.1% · safe right 93.8%

Took 187 of 466 risks that paid · played safe in 541 of 577 positions where safe paid

Error rate 15.4 (95%: 13.1–18.0) · record 4–0–7

League 3

$29.062 per 1,000 decisions$29.062 per 1M decisions (API-equivalent (subscription)) · 1,411 decisions

Parameters: not disclosed by maker

Closed · Flagship ($4 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

DeepSeek-V4.1-Flash · DeepSeek · chat model · low · API + NVIDIA RTX PRO 6000 Blackwell · first release

Risk judgement 69.0 (95%: 67.3–70.9)

Risks found 45.0% · safe right 93.0%

Took 338 of 751 risks that paid · played safe in 891 of 958 positions where safe paid

Error rate 19.0 (95%: 16.6–21.3) · record 12–0–6

League 2

$47.092 per 1,000 decisions$47.092 per 1M decisions (estimate · RTX PRO 6000 Blackwell + API) · 2,226 decisions

Cost: Estimate: $47.092 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 8.2 h on 4 × NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h × 4 = $81.21; plus 1,556 answers from the hosted API at their recorded cost, $23.62; total $104.83 ÷ 2,226 decisions; left out: 38 gaps over 10 min between consecutive answers, 33.3 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 552B backbone + 196B conditional memory; active 8B (prefill) / 16B (decode) · MoE

Mixture of experts: 552B total, about 16B active per step, 8B per token in prefill and 16B in decode (the maker's figures).

Tier (same positions) 1

Released Sep 2026

With its dated checkpoint DeepSeek-V4.1-Flash1017 (its own rows under All setups)

Model page →

Did they improve? Yes · +8.2 pp · threshold 5.0 pp

DeepSeek-V4.1-Flash · DeepSeek · chat model · low · API · 1017

Risk judgement 68.5 (95%: 66.4–70.6)

Risks found 42.5% · safe right 94.5%

Took 321 of 755 risks that paid · played safe in 872 of 923 positions where safe paid

Error rate 17.1 (95%: 15.1–19.2) · record 12–1–5

League 2

$4.869 per 1,000 decisions$4.869 per 1M decisions (API price) · 2,234 decisions

Parameters: not disclosed by maker

Mixture of experts: 552B total, about 16B active per step, 8B per token in prefill and 16B in decode (the maker's figures).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

DeepSeek-V4.1-Flash · DeepSeek · chat model · off · API + NVIDIA RTX PRO 6000 Blackwell · first release

Risk judgement 59.3 (95%: 57.1–61.5)

Risks found 29.1% · safe right 89.5%

Took 192 of 659 risks that paid · played safe in 603 of 674 positions where safe paid

Error rate 35.6 (95%: 31.7–39.7) · record 12–0–6

League 1

$0.728 per 1,000 decisions$0.728 per 1M decisions (estimate · RTX PRO 6000 Blackwell + API) · 1,736 decisions

Cost: Estimate: $0.728 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.07 h on 4 × NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h × 4 = $0.65; plus 1,641 answers from the hosted API at their recorded cost, $0.61; total $1.26 ÷ 1,736 decisions; left out: 7 gaps over 10 min between consecutive answers, 30.8 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 552B backbone + 196B conditional memory; active 8B (prefill) / 16B (decode) · MoE

Mixture of experts: 552B total, about 16B active per step, 8B per token in prefill and 16B in decode (the maker's figures).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

DeepSeek-V4.1-Flash · DeepSeek · chat model · off · API · 1017

Risk judgement 58.5 (95%: 56.6–60.5)

Risks found 25.4% · safe right 91.7%

Took 181 of 713 risks that paid · played safe in 683 of 745 positions where safe paid

Error rate 35.2 (95%: 31.8–38.8) · record 5–1–12

League 2

$0.244 per 1,000 decisions$0.244 per 1M decisions (API price) · 2,034 decisions

Parameters: not disclosed by maker

Mixture of experts: 552B total, about 16B active per step, 8B per token in prefill and 16B in decode (the maker's figures).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

Grok 4.7 · xAI · chat model · high

Risk judgement 68.5 (95%: 66.1–71.1)

Risks found 44.0% · safe right 93.0%

Took 190 of 432 risks that paid · played safe in 454 of 488 positions where safe paid

Error rate 18.8 (95%: 15.9–21.8) · record 6–1–4

League 3

$117.379 per 1,000 decisions$117.379 per 1M decisions (API-equivalent (subscription)) · 1,337 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($2 per 1M input tokens, standard list price).

Tier (same positions) 1

Released Sep 2026

Model page →

Did they improve? pending

Grok 4.7 · xAI · chat model · low

Risk judgement 67.7 (95%: 66.0–69.5)

Risks found 42.3% · safe right 93.0%

Took 295 of 697 risks that paid · played safe in 852 of 916 positions where safe paid

Error rate 19.1 (95%: 16.9–21.4) · record 11–0–7

League 2

$32.813 per 1,000 decisions$32.813 per 1M decisions (API-equivalent (subscription)) · 2,152 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($2 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

Claude Sonnet 5.5 · Anthropic · chat model · high

Risk judgement 68.2 (95%: 66.3–70.0)

Risks found 41.8% · safe right 94.6%

Took 308 of 737 risks that paid · played safe in 841 of 889 positions where safe paid

Error rate 16.7 (95%: 14.9–18.7) · record 12–0–6

League 2

$19.847 per 1,000 decisions$19.847 per 1M decisions (API-equivalent (subscription)) · 2,164 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($2 per 1M input tokens, standard list price).

Tier (same positions) 1

Released Sep 2026

Model page →

Did they improve? Yes · +7.0 pp · threshold 5.0 pp

Claude Sonnet 5.5 · Anthropic · chat model · low

Risk judgement 63.5 (95%: 61.6–65.4)

Risks found 30.9% · safe right 96.2%

Took 204 of 661 risks that paid · played safe in 883 of 918 positions where safe paid

Error rate 20.1 (95%: 17.8–22.4) · record 13–1–4

League 2

$13.541 per 1,000 decisions$13.541 per 1M decisions (API-equivalent (subscription)) · 2,150 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($2 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

Gemini 3.8 Flash · Google · chat model · high

Risk judgement 66.9 (95%: 65.2–68.5)

Risks found 36.1% · safe right 97.6%

Took 264 of 731 risks that paid · played safe in 858 of 879 positions where safe paid

Error rate 14.9 (95%: 12.9–17.0) · record 12–0–6

League 2

$22.352 per 1,000 decisions$22.352 per 1M decisions (API-equivalent (subscription)) · 2,168 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($0.75 per 1M input tokens, standard list price).

Tier (same positions) 1

Released Sep 2026

Model page →

Did they improve? Yes · +8.2 pp · threshold 5.0 pp

Gemini 3.8 Flash · Google · chat model · low

Risk judgement 65.1 (95%: 63.5–66.8)

Risks found 33.8% · safe right 96.5%

Took 214 of 633 risks that paid · played safe in 877 of 909 positions where safe paid

Error rate 16.3 (95%: 14.5–18.2) · record 14–2–2

League 2

$4.123 per 1,000 decisions$4.123 per 1M decisions (API-equivalent (subscription)) · 2,053 decisions

Parameters: not disclosed by maker

Closed · Workhorse ($0.75 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

Muse Ember · Meta · chat model · max

Risk judgement 66.0 (95%: 63.9–68.2)

Risks found 37.6% · safe right 94.4%

Took 224 of 596 risks that paid · played safe in 671 of 711 positions where safe paid

Error rate 20.1 (95%: 17.6–22.7) · record 13–0–4 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$6.863 per 1,000 decisions$6.863 per 1M decisions (API price (billed)) · 1,713 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.1 per 1M input tokens, standard list price).

Tier (same positions) 1

Its late-entry page →

Did they improve? not in the learning plan

Kimi K3.1 Pro · Moonshot · chat model · off

Risk judgement 65.9 (95%: 63.2–68.9)

Risks found 42.4% · safe right 89.5%

Took 153 of 361 risks that paid · played safe in 416 of 465 positions where safe paid

Error rate 28.0 (95%: 24.5–31.8) · record 4–0–7

League 3

$3.982 per 1,000 decisions$3.982 per 1M decisions (API price) · 1,167 decisions

Parameters: 2.8T MoE · 104B active

Mixture of experts: 2.8T total, about 104B active per step (the maker's figures).

Tier (same positions) 1

Released Oct 2026

Model page →

Did they improve? No · +4.8 pp · threshold 5.0 pp

Gemini Decision 1 · Google · decision model · 20261111

Risk judgement 64.7 (95%: 62.2–67.2)

Risks found 40.9% · safe right 88.5%

Took 219 of 535 risks that paid · played safe in 648 of 732 positions where safe paid

Error rate 28.0 (95%: 25.0–31.1) · record 13–0–4 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.382 per 1,000 decisions$0.382 per 1M decisions (API price (billed)) · 1,704 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.1 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

Gemma 4 31B · Google · chat model · on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 64.5 (95%: 62.9–66.2)

Risks found 31.9% · safe right 97.1%

Took 223 of 700 risks that paid · played safe in 878 of 904 positions where safe paid

Error rate 19.6 (95%: 17.5–21.7) · record 13–0–5

League 2

$34.943 per 1,000 decisions$34.943 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,122 decisions

Cost: Estimate: $34.943 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.3 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.27 + 22.1 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $55.13 + 3.8 h on 2 × NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h × 2 = $18.75; total $74.15 ÷ 2,122 decisions; left out: 62 gaps over 10 min between consecutive answers, 134.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 31.27B incl. its vision encoder · dense

Size class by the maker's figure, 30.7B; our checkpoint count 31.27B.

Tier (same positions) 1

Released Apr 2026

Model page →

Did they improve? Yes · +8.5 pp · threshold 5.0 pp

Gemma 4 31B · Google · chat model · off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 62.6 (95%: 60.4–65.0)

Risks found 33.2% · safe right 92.1%

Took 194 of 584 risks that paid · played safe in 732 of 795 positions where safe paid

Error rate 22.5 (95%: 19.9–25.3) · record 10–0–8

League 2

$8.215 per 1,000 decisions$8.215 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 1,847 decisions

Cost: Estimate: $8.215 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.34 + 6.0 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $14.83; total $15.17 ÷ 1,847 decisions; left out: 10 gaps over 10 min between consecutive answers, 151.2 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 31.27B incl. its vision encoder · dense

Size class by the maker's figure, 30.7B; our checkpoint count 31.27B.

Tier (same positions) not run yet

Released Apr 2026

Model page →

Did they improve? not in the learning plan

d1-ultra · Liquid AI · decision model

Risk judgement 64.0 (95%: 61.7–66.3)

Risks found 42.0% · safe right 86.0%

Took 289 of 688 risks that paid · played safe in 702 of 816 positions where safe paid

Error rate 30.3 (95%: 27.3–33.4) · record 6–1–11

League 2

$0.145 per 1,000 decisions$0.145 per 1M decisions (API price) · 2,071 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.042 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? (decision models, two arms) not tested (conservative context guard)

Mammoth-1 · Base Labs · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 63.2 (95%: 61.2–65.2)

Risks found 40.0% · safe right 86.4%

Took 288 of 720 risks that paid · played safe in 760 of 880 positions where safe paid

Error rate 28.5 (95%: 26.2–31.1) · record 16–1–2 · late entry

Played the original League 1 field after the league finished (late entry)

$0.999 per 1,000 decisions$0.999 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,021 decisions

Cost: Estimate: $0.999 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.8 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $2.02; total $2.02 ÷ 2,021 decisions; left out: 1 gaps over 10 min between consecutive answers, 0.7 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 27.36B incl. its vision encoder; card: 27B · dense

Size class by the maker's figure, 27B; our checkpoint count 27.36B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) waiting (Ümit's D5 decision)

Drex Max · Nace.AI · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 62.9 (95%: 61.1–64.7)

Risks found 39.3% · safe right 86.5%

Took 299 of 761 risks that paid · played safe in 775 of 896 positions where safe paid

Error rate 30.4 (95%: 27.4–33.4) · record 11–0–7 · late entry

Played the original League 1 field after the league finished (late entry)

18 of 19 matches · GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)

$1.285 per 1,000 decisions$1.285 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,148 decisions

Cost: Estimate: $1.285 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.1 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $2.76; total $2.76 ÷ 2,148 decisions; left out: 1 gaps over 10 min between consecutive answers, 5.3 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: backbone 26.09B incl. its vision encoder + decision head 1.3M; card: 27B · dense

Size class by the maker's figure, 27B; our checkpoint count 26.09B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

Sieve 31B v3 · Sthānika AI · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 62.7 (95%: 60.6–64.9)

Risks found 35.2% · safe right 90.2%

Took 200 of 568 risks that paid · played safe in 696 of 772 positions where safe paid

Error rate 27.5 (95%: 24.1–31.6) · record 15–0–2 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.794 per 1,000 decisions$0.794 per 1M decisions (estimate · DGX Spark) · 1,740 decisions

Cost: Estimate: $0.794 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 6.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $1.38; total $1.38 ÷ 1,740 decisions; left out: 1 gaps over 10 min between consecutive answers, 32.5 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 31.27B incl. its vision encoder; card: 31B (backbone gemma-4-31B-it; 29.29B trained) · dense

Size class by the maker's figure, 31B; our checkpoint count 31.27B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

Qwen3.8-27B · Alibaba · chat model · low · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 62.0 (95%: 60.1–63.9)

Risks found 33.8% · safe right 90.2%

Took 242 of 715 risks that paid · played safe in 735 of 815 positions where safe paid

Error rate 29.0 (95%: 26.5–31.8) · record 9–0–9

League 2

$39.023 per 1,000 decisions$39.023 per 1M decisions (estimate · RTX PRO 6000 Blackwell + API) · 2,108 decisions

Cost: Estimate: $39.023 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 5.0 h on NVIDIA DGX Spark (GB10) × $0.213/h = $1.07 + 32.6 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $81.19; plus 1 answers from the hosted API at their recorded cost, $0.00; total $82.26 ÷ 2,108 decisions; left out: 61 gaps over 10 min between consecutive answers, 122.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 27.78B incl. its vision encoder; card: 27B · dense

Size class by the maker's figure, 27B; our checkpoint count 27.78B.

Tier (same positions) 2

Released Aug 2026

Model page →

Did they improve? No · +19.0 pp · threshold 5.0 pp

Qwen3.8-27B · Alibaba · chat model · off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 57.6 (95%: 56.1–59.0)

Risks found 23.5% · safe right 91.7%

Took 178 of 756 risks that paid · played safe in 671 of 732 positions where safe paid

Error rate 45.0 (95%: 40.7–49.9) · record 9–1–8

League 1

$4.551 per 1,000 decisions$4.551 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,005 decisions

Cost: Estimate: $4.551 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.06 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.01 + 3.7 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $9.11; total $9.13 ÷ 2,005 decisions; left out: 6 gaps over 10 min between consecutive answers, 117.9 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 27.78B incl. its vision encoder; card: 27B · dense

Size class by the maker's figure, 27B; our checkpoint count 27.78B.

Tier (same positions) not run yet

Released Aug 2026

Model page →

Did they improve? not in the learning plan

GLM-5.3 · Z.ai · chat model · low

Risk judgement 61.7 (95%: 59.7–63.8)

Risks found 30.7% · safe right 92.7%

Took 205 of 668 risks that paid · played safe in 726 of 783 positions where safe paid

Error rate 30.9 (95%: 27.8–34.0) · record 7–0–11

League 2

$2.917 per 1,000 decisions$2.917 per 1M decisions (API price) · 1,989 decisions

Parameters: 753.33B MoE; active not disclosed by maker

The maker's model card states no total; 753.33B is counted from the released files (Hugging Face's file total agrees).

Tier (same positions) 1

Released Aug 2026

Model page →

Did they improve? Yes · +17.6 pp · threshold 9.0 pp

KnowLine-31B-Gen1.5 · PelaAI · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 61.6 (95%: 59.6–63.5)

Risks found 28.8% · safe right 94.4%

Took 187 of 649 risks that paid · played safe in 758 of 803 positions where safe paid

Error rate 26.2 (95%: 23.0–29.4) · record 13–1–3 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

1,928 decisions

Parameters: 31.27B incl. its vision encoder; card: about 31B · dense

Size class by the maker's figure, 31B; our checkpoint count 31.27B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

REAX 1.3 · REAX · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 60.5 (95%: 58.7–62.4)

Risks found 31.7% · safe right 89.4%

Took 211 of 666 risks that paid · played safe in 760 of 850 positions where safe paid

Error rate 27.8 (95%: 24.9–30.6) · record 11–1–6 · late entry

Played the original League 1 field after the league finished (late entry)

18 of 19 matches · GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)

$1.545 per 1,000 decisions$1.545 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 1,998 decisions

Cost: Estimate: $1.545 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.2 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $3.09; total $3.09 ÷ 1,998 decisions; left out: 2 gaps over 10 min between consecutive answers, 4.9 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: part G 31.27B + part Q 26.09B (each incl. its vision encoder) + decision head 1.3M · dense

Two models run together, option probabilities averaged 50/50: 57.36B in all, all used on every question.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) waiting (Ümit's D5 decision)

GLiDE · Fastino · decision model · API

Risk judgement 60.4 (95%: 58.3–62.6)

Risks found 36.9% · safe right 84.0%

Took 249 of 675 risks that paid · played safe in 756 of 900 positions where safe paid

Error rate 33.2 (95%: 30.2–36.4) · record 5–0–13

League 2

$0.153 per 1,000 decisions$0.153 per 1M decisions (API price) · 2,159 decisions

Cost: List price × our recorded tokens: 8,267,523 input tokens × $0.04 per 1M (output $0) = $0.33 ÷ 2,159 decisions (https://docs.fastino.ai/pricing)

Parameters: not disclosed by maker

Closed · Efficient ($0.04 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

Turion Jev 31B · Turion.AI · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 59.6 (95%: 57.7–61.4)

Risks found 24.0% · safe right 95.1%

Took 162 of 674 risks that paid · played safe in 680 of 715 positions where safe paid

Error rate 29.7 (95%: 26.5–33.3) · record 13–1–3 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.293 per 1,000 decisions$0.293 per 1M decisions (estimate · DGX Spark) · 1,855 decisions

Cost: Estimate: $0.293 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 2.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.54; total $0.54 ÷ 1,855 decisions; left out: 1 gaps over 10 min between consecutive answers, 14.6 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 31.27B incl. its vision encoder; card: 31B (base gemma-4-31B-it) · dense

Size class by the maker's figure, 31B; our checkpoint count 31.27B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

Jev 1.13 · TypeSafe · decision model

Risk judgement 59.5 (95%: 57.4–61.6)

Risks found 27.0% · safe right 92.1%

Took 190 of 705 risks that paid · played safe in 710 of 771 positions where safe paid

Error rate 34.4 (95%: 31.0–38.0) · record 12–0–6

League 1

$0.166 per 1,000 decisions$0.166 per 1M decisions (API price) · 1,911 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.042 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Model page →

Did they improve? not in the learning plan

Qwen3.8 Flash · Alibaba · chat model · off

Risk judgement 59.5 (95%: 57.2–61.7)

Risks found 34.4% · safe right 84.6%

Took 243 of 706 risks that paid · played safe in 613 of 725 positions where safe paid

Error rate 33.6 (95%: 30.8–36.7) · record 2–0–16

League 2

$0.284 per 1,000 decisions$0.284 per 1M decisions (API price) · 2,027 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.15 per 1M input tokens, standard list price).

Tier (same positions) 2

Model page →

Did they improve? No · +2.7 pp · threshold 5.0 pp

Jev 1.14 · TypeSafe · decision model

Risk judgement 59.3 (95%: 57.1–61.4)

Risks found 29.1% · safe right 89.4%

Took 184 of 632 risks that paid · played safe in 657 of 735 positions where safe paid

Error rate 31.7 (95%: 28.3–35.7) · record 5–0–13

League 2

$0.173 per 1,000 decisions$0.173 per 1M decisions (API price) · 1,912 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.042 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? (decision models, two arms) not tested (conservative context guard)

Drex Max v1.1 · Nace.AI · decision model

Risk judgement 59.1 (95%: 57.0–61.3)

Risks found 35.6% · safe right 82.5%

Took 250 of 702 risks that paid · played safe in 775 of 939 positions where safe paid

Error rate 30.9 (95%: 28.2–33.7) · record 12–2–4 · late entry

Played the original League 1 field after the league finished (late entry)

18 of 19 matches · GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)

$0.265 per 1,000 decisions$0.265 per 1M decisions (list price; output not billed) · 2,154 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.03 per 1M input tokens, introductory price, the only one published); class provisional: Only its launch price is published: the models reference lists $0.03, and nace.ai/developers says 'Drex Max v1.1 cost is at its 50% launch price'; no regular price for v1.1 is printed (Drex MaxNVIDIA RTX PRO 6000 Blackwell's launch post gave $0.06 for the earlier model, not used).

Tier (same positions) 2

Its late-entry page →

Did they improve? (decision models, two arms) notes: complete · worst: waiting (cash approval)

GLiDE · Fastino · decision model · v1.2 · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 58.8 (95%: 56.7–60.8)

Risks found 33.0% · safe right 84.6%

Took 219 of 664 risks that paid · played safe in 698 of 825 positions where safe paid

Error rate 35.0 (95%: 31.6–39.0) · record 12–0–6

League 1

$10.301 per 1,000 decisions$10.301 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,003 decisions

Cost: Estimate: $10.301 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 8.3 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $20.63; total $20.63 ÷ 2,003 decisions; left out: 26 gaps over 10 min between consecutive answers, 81.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: backbone 26.09B incl. its vision encoder + decision head 1.3M · dense

Size class by the maker's figure, 27B; our checkpoint count 26.09B.

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? (decision models, two arms) notes: stopped (its budget cap was reached) · worst: complete

Drex 2.0 · Nace.AI · decision model · API

Risk judgement 58.1 (95%: 55.7–60.4)

Risks found 31.5% · safe right 84.8%

Took 220 of 699 risks that paid · played safe in 623 of 735 positions where safe paid

Error rate 38.4 (95%: 35.6–41.4) · record 10–0–8 · late entry

Played the original League 1 field after the league finished (late entry)

18 of 19 matches · GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)

$0.258 per 1,000 decisions$0.258 per 1M decisions (list price (billed); output not billed) · 1,937 decisions

Parameters: 35.11B incl. a vision tower the card says is dropped on load; card: 35B; 3B active · MoE

Mixture of experts: 35B total, about 3B active per step (the maker's figures); our checkpoint count 35.11B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

Drex 2.0 · Nace.AI · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 53.0 (95%: 50.5–55.2)

Risks found 26.2% · safe right 79.7%

Took 189 of 720 risks that paid · played safe in 581 of 729 positions where safe paid

Error rate 42.9 (95%: 39.1–47.1) · record 7–0–11 · late entry

Played the original League 1 field after the league finished (late entry)

18 of 19 matches · GPT-6 LunaDecisions API: outside the requested scope (Ümit, 12 Nov); not played (opponent's API out of credit)

$0.782 per 1,000 decisions$0.782 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 1,993 decisions

Cost: Estimate: $0.782 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.6 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $1.56; total $1.56 ÷ 1,993 decisions; no gap over 10 min. NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 35.11B incl. a vision tower the card says is dropped on load; card: 35B; 3B active · MoE

Mixture of experts: 35B total, about 3B active per step (the maker's figures); our checkpoint count 35.11B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

Perplexity Decider 27B v1.2 · Perplexity · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 57.9 (95%: 56.0–59.9)

Risks found 33.4% · safe right 82.5%

Took 244 of 731 risks that paid · played safe in 682 of 827 positions where safe paid

Error rate 36.9 (95%: 33.9–40.1) · record 12–0–6

League 1

$12.227 per 1,000 decisions$12.227 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,028 decisions

Cost: Estimate: $12.227 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 10.0 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $24.80; total $24.80 ÷ 2,028 decisions; left out: 19 gaps over 10 min between consecutive answers, 131.1 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: backbone 26.09B incl. its vision encoder + decision head 1.3M; changelog: 27B · dense

Size class by the maker's figure, 27B; our checkpoint count 26.09B.

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? (decision models, two arms) complete

celeris-1-decision · Celeris · decision model

Risk judgement 57.8 (95%: 55.9–59.7)

Risks found 23.9% · safe right 91.7%

Took 163 of 683 risks that paid · played safe in 749 of 817 positions where safe paid

Error rate 35.6 (95%: 32.0–39.4) · record 12–2–4

League 1

$0.150 per 1,000 decisions$0.150 per 1M decisions (API price) · 1,922 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.04 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? not in the learning plan

Elephant-1 · Base Labs · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 57.7 (95%: 55.8–59.6)

Risks found 28.4% · safe right 87.0%

Took 231 of 812 risks that paid · played safe in 682 of 784 positions where safe paid

Error rate 44.6 (95%: 40.3–48.9) · record 11–0–8 · late entry

Played the original League 1 field after the league finished (late entry)

$0.685 per 1,000 decisions$0.685 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,044 decisions

Cost: Estimate: $0.685 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.6 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $1.40; total $1.40 ÷ 2,044 decisions; left out: 2 gaps over 10 min between consecutive answers, 1.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 9.41B incl. its vision encoder; card: 9B · dense

Size class by the maker's figure, 9B; our checkpoint count 9.41B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (Ümit's D5 decision) · worst: complete

Clef · Cloudflare · decision model · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 57.5 (95%: 55.4–59.8)

Risks found 39.5% · safe right 75.6%

Took 263 of 666 risks that paid · played safe in 603 of 798 positions where safe paid

Error rate 36.5 (95%: 33.5–39.6) · record 2–0–16

League 2

$5.736 per 1,000 decisions$5.736 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 1,978 decisions

Cost: Estimate: $5.736 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.34 + 4.4 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $11.01; total $11.35 ÷ 1,978 decisions; left out: 7 gaps over 10 min between consecutive answers, 243.2 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: backbone 27.36B incl. its vision encoder + joint head 128.1M; card: 27B · dense

Size class by the maker's figure, 27B; our checkpoint count 27.48B.

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? not in the learning plan

StartLux-Decision-v1.1-4B-Instant · StartLux · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 57.2 (95%: 55.0–59.3)

Risks found 34.1% · safe right 80.4%

Took 265 of 777 risks that paid · played safe in 594 of 739 positions where safe paid

Error rate 40.1 (95%: 36.7–43.7) · record 10–0–7 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.072 per 1,000 decisions$0.072 per 1M decisions (estimate · DGX Spark) · 1,989 decisions

Cost: Estimate: $0.072 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.7 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.14; total $0.14 ÷ 1,989 decisions; left out: 1 gaps over 10 min between consecutive answers, 12.5 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 4.66B incl. its vision encoder and MTP layer; card: 4B · dense

Size class by the maker's figure, 4B; our checkpoint count 4.66B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

Cygnet · Blockbrain Labs · decision model · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 57.1 (95%: 55.3–58.8)

Risks found 27.3% · safe right 86.8%

Took 205 of 750 risks that paid · played safe in 685 of 789 positions where safe paid

Error rate 38.6 (95%: 35.4–41.9) · record 12–1–5

League 1

$11.168 per 1,000 decisions$11.168 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,016 decisions

Cost: Estimate: $11.168 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.12 + 9.0 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $22.39; total $22.51 ÷ 2,016 decisions; left out: 19 gaps over 10 min between consecutive answers, 252.8 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 11.96B incl. its vision encoder; card: 11.95B · dense

Size class by the maker's figure, 11.95B; our checkpoint count 11.96B.

Base model: Gemma 4 12B IT

Tier (same positions) not run yet

Model page →

Did they improve? (decision models, two arms) waiting (its layer never sees the lessons)

Perplexity Decider 27B v1.1 · Perplexity · decision model · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 56.9 (95%: 54.9–58.8)

Risks found 33.1% · safe right 80.7%

Took 233 of 703 risks that paid · played safe in 682 of 845 positions where safe paid

Error rate 38.8 (95%: 35.3–42.5) · record 10–0–8

League 1

$6.593 per 1,000 decisions$6.593 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,052 decisions

Cost: Estimate: $6.593 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 2.3 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.49 + 5.2 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $13.04; total $13.53 ÷ 2,052 decisions; left out: 34 gaps over 10 min between consecutive answers, 255.1 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: backbone 26.09B incl. its vision encoder + decision head 1.3M · dense

Size class by the maker's figure, 27B; our checkpoint count 26.09B.

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? not in the learning plan

SubQ Snap 1.1 · Subquadratic · decision model

Risk judgement 56.7 (95%: 54.8–58.6)

Risks found 38.6% · safe right 74.8%

Took 308 of 798 risks that paid · played safe in 750 of 1,003 positions where safe paid

Error rate 40.2 (95%: 36.7–43.8) · record 12–0–7 · late entry

Played the original League 1 field after the league finished (late entry)

This 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).

$0.150 per 1,000 decisions$0.150 per 1M decisions (API price (billed); output free) · 2,326 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.04 per 1M input tokens, standard list price).

Tier (same positions) 2

Its late-entry page →

Did they improve? (decision models, two arms) complete

SubQ Snap Preview · Subquadratic · decision model

Risk judgement 56.5 (95%: 53.8–59.1)

Risks found 37.9% · safe right 75.1%

Took 297 of 783 risks that paid · played safe in 673 of 896 positions where safe paid

Error rate 43.9 (95%: 39.3–48.8) · record 12–1–5

League 1

$0.193 per 1,000 decisions$0.193 per 1M decisions (API price) · 2,200 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.05 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Model page →

Did they improve? not in the learning plan

Cerno 9B v1.2 · Apertix AI · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 56.1 (95%: 53.9–58.2)

Risks found 39.2% · safe right 72.9%

Took 288 of 734 risks that paid · played safe in 625 of 857 positions where safe paid

Error rate 48.5 (95%: 44.5–52.6) · record 6–1–10 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.139 per 1,000 decisions$0.139 per 1M decisions (estimate · DGX Spark) · 2,108 decisions

Cost: Estimate: $0.139 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.29; total $0.29 ÷ 2,108 decisions; left out: 1 gaps over 10 min between consecutive answers, 11.5 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 8.95B (untied lm_head); card: 9B · dense

Size class by the maker's figure, 9B; our checkpoint count 8.95B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

GPT-6 Luna · OpenAI · decision model · Decisions API

Risk judgement 55.7 (95%: 53.8–57.6)

Risks found 21.6% · safe right 89.7%

Took 160 of 740 risks that paid · played safe in 636 of 709 positions where safe paid

Error rate 43.3 (95%: 39.7–46.9) · record 11–1–6

League 1

$0.335 per 1,000 decisions$0.335 per 1M decisions (API price) · 1,962 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.1 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Released Sep 2026

Model page →

Did they improve? not in the learning plan

d1-2B · Liquid AI · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 55.2 (95%: 52.8–57.5)

Risks found 33.6% · safe right 76.8%

Took 231 of 688 risks that paid · played safe in 596 of 776 positions where safe paid

Error rate 47.1 (95%: 42.8–51.5) · record 5–0–12 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.467 per 1,000 decisions$0.467 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 1,980 decisions

Cost: Estimate: $0.467 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $0.93; total $0.93 ÷ 1,980 decisions; left out: 1 gaps over 10 min between consecutive answers, 2.5 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 1.92B incl. its vision tower; card: 1.92B · dense

Size class by the maker's figure, 1.92B.

Base model: LFM3-VL-2B

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) not started (no backgammon route yet; no league record)

Winnow-12B · EldanRing · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 55.2 (95%: 52.9–57.2)

Risks found 33.7% · safe right 76.6%

Took 263 of 780 risks that paid · played safe in 573 of 748 positions where safe paid

Error rate 49.7 (95%: 45.9–53.6) · record 7–0–11

League 1

$9.615 per 1,000 decisions$9.615 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,015 decisions

Cost: Estimate: $9.615 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 7.8 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $19.37; total $19.37 ÷ 2,015 decisions; left out: 22 gaps over 10 min between consecutive answers, 149.6 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 11.91B dense

Size class by the maker's figure, 11.95B; our checkpoint count 11.91B.

Tier (same positions) not run yet

Model page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

StartLux-Decision-v1.1-2B-Instant · StartLux · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 53.8 (95%: 51.3–56.6)

Risks found 33.7% · safe right 74.0%

Took 212 of 630 risks that paid · played safe in 450 of 608 positions where safe paid

Error rate 50.0 (95%: 45.4–54.8) · record 3–0–14 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.052 per 1,000 decisions$0.052 per 1M decisions (estimate · DGX Spark) · 1,771 decisions

Cost: Estimate: $0.052 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.09; total $0.09 ÷ 1,771 decisions; left out: 1 gaps over 10 min between consecutive answers, 63.8 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 2.27B incl. its vision encoder and MTP layer; card: 2B · dense

Size class by the maker's figure, 2B; our checkpoint count 2.27B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

Mercury Decide · Inception · decision model

Risk judgement 53.4 (95%: 51.5–55.2)

Risks found 18.7% · safe right 88.1%

Took 129 of 690 risks that paid · played safe in 583 of 662 positions where safe paid

Error rate 47.9 (95%: 43.3–52.9) · record 8–1–9

League 1

$0.071 per 1,000 decisions$0.071 per 1M decisions (API price) · 1,853 decisions

Parameters: not disclosed by maker

Closed · Efficient ($0.04 per 1M input tokens, standard list price).

Tier (same positions) not run yet

Model page →

Did they improve? not in the learning plan

d1-Omni-2B · Liquid AI · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 53.2 (95%: 50.9–55.6)

Risks found 30.4% · safe right 76.0%

Took 188 of 619 risks that paid · played safe in 553 of 728 positions where safe paid

Error rate 49.8 (95%: 43.5–57.4) · record 9–0–8 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.657 per 1,000 decisions$0.657 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 1,710 decisions

Cost: Estimate: $0.657 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $1.12; total $1.12 ÷ 1,710 decisions; no gap over 10 min. NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 2.17B incl. its vision and audio parts; card: 2.17B · dense

Size class by the maker's figure, 2.17B.

Base model: LFM3-Omni-2B

Tier (same positions) not run yet

Its late-entry page →

Did they improve? not in the learning plan

LFM2.5-24B-A2B · Liquid AI · chat model · on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark

Risk judgement 52.4 (95%: 50.3–54.5)

Risks found 31.5% · safe right 73.4%

Took 248 of 788 risks that paid · played safe in 647 of 882 positions where safe paid

Error rate 47.7 (95%: 44.1–51.2) · record 10–0–8

League 1

$38.716 per 1,000 decisions$38.716 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,193 decisions

Cost: Estimate: $38.716 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.1 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.02 + 34.1 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $84.88; total $84.90 ÷ 2,193 decisions; left out: 16 gaps over 10 min between consecutive answers, 97.1 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 23.97B MoE · 2.45B active

Mixture of experts: 24B total, about 2.45B active per step (the maker's figures); our checkpoint count 23.97B.

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? pending

LFM2.5-24B-A2B · Liquid AI · chat model · off · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 45.7 (95%: 43.2–48.1)

Risks found 28.8% · safe right 62.7%

Took 206 of 715 risks that paid · played safe in 470 of 750 positions where safe paid

Error rate 61.5 (95%: 56.3–66.8) · record 3–0–15

League 1

$9.132 per 1,000 decisions$9.132 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 1,952 decisions

Cost: Estimate: $9.132 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 7.2 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $17.83; total $17.83 ÷ 1,952 decisions; left out: 19 gaps over 10 min between consecutive answers, 87.9 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 23.97B MoE · 2.45B active

Mixture of experts: 24B total, about 2.45B active per step (the maker's figures); our checkpoint count 23.97B.

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? not in the learning plan

KnowLine-2B-Gen1 · PelaAI · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 51.1 (95%: 48.7–53.5)

Risks found 30.0% · safe right 72.3%

Took 203 of 676 risks that paid · played safe in 508 of 703 positions where safe paid

Error rate 54.5 (95%: 48.9–60.8) · record 3–0–14 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.049 per 1,000 decisions$0.049 per 1M decisions (estimate · DGX Spark) · 1,955 decisions

Cost: Estimate: $0.049 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.10; total $0.10 ÷ 1,955 decisions; left out: 1 gaps over 10 min between consecutive answers, 34.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 2.27B incl. its vision encoder and MTP layer; card: 2B · dense

Size class by the maker's figure, 2B; our checkpoint count 2.27B.

Base model: Qwen3.5-2B

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

KnowLine 4B Gen5.5 · PelaAI · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 49.7 (95%: 47.0–52.3)

Risks found 34.1% · safe right 65.4%

Took 265 of 778 risks that paid · played safe in 532 of 813 positions where safe paid

Error rate 53.6 (95%: 48.8–58.6) · record 5–1–11 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.054 per 1,000 decisions$0.054 per 1M decisions (estimate · DGX Spark) · 2,153 decisions

Cost: Estimate: $0.054 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.12; total $0.12 ÷ 2,153 decisions; left out: 1 gaps over 10 min between consecutive answers, 64.1 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 4.66B incl. its vision encoder and MTP layer; card: 4B · dense

Size class by the maker's figure, 4B; our checkpoint count 4.66B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

Perplexity Decider Flash v1.2 · Perplexity · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 49.1 (95%: 46.5–51.6)

Risks found 31.3% · safe right 66.9%

Took 223 of 712 risks that paid · played safe in 519 of 776 positions where safe paid

Error rate 59.6 (95%: 54.9–64.5) · record 2–0–16

League 1

$12.267 per 1,000 decisions$12.267 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,006 decisions

Cost: Estimate: $12.267 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 9.9 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $24.61; total $24.61 ÷ 2,006 decisions; left out: 22 gaps over 10 min between consecutive answers, 147.3 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: backbone 4.54B incl. its vision encoder + decision head 0.7M; changelog: 4B · dense

Size class by the maker's figure, 4B; our checkpoint count 4.54B.

Tier (same positions) not run yet

Released Oct 2026

Model page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

Open-Jev-27B-v1.1 · Zefan Cai · decision model · NVIDIA RTX PRO 6000 Blackwell

Risk judgement 48.5 (95%: 46.1–50.9)

Risks found 37.3% · safe right 59.8%

Took 264 of 708 risks that paid · played safe in 455 of 761 positions where safe paid

Error rate 65.6 (95%: 59.7–71.9) · record 2–1–15

League 1

$7.185 per 1,000 decisions$7.185 per 1M decisions (estimate · RTX PRO 6000 Blackwell) · 2,002 decisions

Cost: Estimate: $7.185 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 5.8 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $14.38; total $14.38 ÷ 2,002 decisions; left out: 13 gaps over 10 min between consecutive answers, 257.6 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 27.80B: Qwen3.8-27B base 27.78B + adapter and head 15.47M (the repo holds only the adapter and head) · dense

Size class by the maker's figure, 27B; our checkpoint count 27.8B.

Base model: Qwen3.8-27B + adapter

Tier (same positions) not run yet

Model page →

Did they improve? not in the learning plan

GLiDE 2B · Fastino · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 46.7 (95%: 43.6–49.6)

Risks found 33.0% · safe right 60.3%

Took 192 of 581 risks that paid · played safe in 354 of 587 positions where safe paid

Error rate 67.4 (95%: 61.0–74.5) · record 1–0–16 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.058 per 1,000 decisions$0.058 per 1M decisions (estimate · DGX Spark) · 1,676 decisions

Cost: Estimate: $0.058 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.10; total $0.10 ÷ 1,676 decisions; left out: 1 gaps over 10 min between consecutive answers, 30.9 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 1.88B incl. its decision readout; card: 1.88B · dense

Size class by the maker's figure, 1.88B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

KnowLine-0.8B-Gen1 · PelaAI · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 45.2 (95%: 42.5–47.7)

Risks found 30.8% · safe right 59.6%

Took 196 of 637 risks that paid · played safe in 354 of 594 positions where safe paid

Error rate 71.9 (95%: 66.2–77.9) · record 1–1–15 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.049 per 1,000 decisions$0.049 per 1M decisions (estimate · DGX Spark) · 1,728 decisions

Cost: Estimate: $0.049 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.08; total $0.08 ÷ 1,728 decisions; left out: 1 gaps over 10 min between consecutive answers, 55.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 0.87B incl. its vision encoder and MTP layer; card: 0.8B · dense

Size class by the maker's figure, 0.8B; our checkpoint count 0.87B.

Base model: Qwen3.5-0.8B

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

StartLux-Decision-v1.1-0.8B-Instant · StartLux · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 44.6 (95%: 42.3–46.9)

Risks found 38.3% · safe right 50.9%

Took 255 of 665 risks that paid · played safe in 364 of 715 positions where safe paid

Error rate 64.7 (95%: 59.3–70.3) · record 3–0–14 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.054 per 1,000 decisions$0.054 per 1M decisions (estimate · DGX Spark) · 1,882 decisions

Cost: Estimate: $0.054 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.10; total $0.10 ÷ 1,882 decisions; left out: 1 gaps over 10 min between consecutive answers, 54.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 0.87B incl. its vision encoder and MTP layer; card: 0.8B · dense

Size class by the maker's figure, 0.8B; our checkpoint count 0.87B.

Base model: Qwen3.5-0.8B

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

OneJev1.6-0.8B · OmniJev · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 42.5 (95%: 39.3–45.8)

Risks found 27.6% · safe right 57.4%

Took 167 of 606 risks that paid · played safe in 364 of 634 positions where safe paid

Error rate 82.5 (95%: 75.2–90.8) · record 0–0–17 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.052 per 1,000 decisions$0.052 per 1M decisions (estimate · DGX Spark) · 1,763 decisions

Cost: Estimate: $0.052 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.09; total $0.09 ÷ 1,763 decisions; left out: 1 gaps over 10 min between consecutive answers, 53.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 0.85B incl. its vision encoder; card: 0.9B · dense

Size class by the maker's figure, 0.9B; our checkpoint count 0.85B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

GLiDE 0.8B · Fastino · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 41.7 (95%: 38.9–44.3)

Risks found 30.4% · safe right 53.1%

Took 204 of 671 risks that paid · played safe in 352 of 663 positions where safe paid

Error rate 80.4 (95%: 71.8–89.3) · record 2–1–14 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.051 per 1,000 decisions$0.051 per 1M decisions (estimate · DGX Spark) · 1,832 decisions

Cost: Estimate: $0.051 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.09; total $0.09 ÷ 1,832 decisions; left out: 1 gaps over 10 min between consecutive answers, 31.4 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 0.75B incl. its decision readout; card: 0.75B · dense

Size class by the maker's figure, 0.75B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

GLiDE 0.4B · Fastino · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 28.9 (95%: 25.6–32.1)

Risks found 18.8% · safe right 39.0%

Took 83 of 441 risks that paid · played safe in 194 of 498 positions where safe paid

Error rate 113.2 (95%: 102.1–124.1) · record 1–0–16 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.057 per 1,000 decisions$0.057 per 1M decisions (estimate · DGX Spark) · 1,355 decisions

Cost: Estimate: $0.057 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.08; total $0.08 ÷ 1,355 decisions; left out: 1 gaps over 10 min between consecutive answers, 31.7 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 0.40B incl. its decision readout; card: 0.4B · dense

Size class by the maker's figure, 0.4B.

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

GLiClass v4 Edge · Knowledgator · decision model · NVIDIA DGX Spark (GB10)

Risk judgement 21.5 (95%: 18.6–24.6)

Risks found 26.8% · safe right 16.3%

Took 136 of 508 risks that paid · played safe in 78 of 478 positions where safe paid

Error rate 133.5 (95%: 118.5–148.8) · record 0–0–17 · late entry

Played the original League 1 field after the league finished (late entry)

17 of 19 matches · GPT-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)

$0.174 per 1,000 decisions$0.174 per 1M decisions (estimate · DGX Spark) · 1,405 decisions

Cost: Estimate: $0.174 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.2 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.25; total $0.25 ÷ 1,405 decisions; left out: 2 gaps over 10 min between consecutive answers, 42.1 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Parameters: 0.44B incl. its projectors, scorer and calibrator; card: 0.4B · dense

Size class by the maker's figure, 0.4B; our checkpoint count 0.44B.

Base model: DeBERTa-v3-large

Tier (same positions) not run yet

Its late-entry page →

Did they improve? (decision models, two arms) notes: waiting (cash approval) · worst: waiting (its render check)

GNUbg weak · x86-64 build: its noise is shifted, so it overplays gammons

Risk judgement 67.1 (95%: 63.8–70.3)

Risks found 66.6% · safe right 67.6%

Took 410 of 616 risks that paid · played safe in 533 of 788 positions where safe paid

Error rate 21.4 (95%: 18.8–24.3) · record 16–0–2

League 1

1,744 decisions

Model page →

GNUbg weak · x86-64 build: its noise is shifted, so it overplays gammons

Risk judgement 62.3 (95%: 59.5–64.8)

Risks found 62.7% · safe right 61.9%

Took 483 of 770 risks that paid · played safe in 558 of 902 positions where safe paid

Error rate 27.0 (95%: 24.1–29.8) · record 8–0–10

League 2

2,136 decisions

Model page →

GNUbg weak · x86-64 build: its noise is shifted, so it overplays gammons

Risk judgement 62.1 (95%: 58.2–65.9)

Risks found 58.8% · safe right 65.5%

Took 288 of 490 risks that paid · played safe in 415 of 634 positions where safe paid

Error rate 28.1 (95%: 25.0–31.4) · record 2–2–7

League 3

1,479 decisions

Model page →

Safest · always plays safe

Risk judgement 50.0 (95%: 50.0–50.0)

Risks found 0.0% · safe right 100.0%

Took 0 of 785 risks that paid · played safe in 749 of 749 positions where safe paid

Error rate 51.6 (95%: 47.4–56.2) · record 4–2–12

League 1

2,058 decisions

Model page →

Safest (computed)

Risk judgement 50.0 (95%: 50.0–50.0)

Risks found 0.0% · safe right 100.0%

Took 0 of 13,140 risks that paid · played safe in 16,136 of 16,136 positions where safe paid

League 2

Safest (computed)

Risk judgement 50.0 (95%: 50.0–50.0)

Risks found 0.0% · safe right 100.0%

Took 0 of 5,153 risks that paid · played safe in 6,771 of 6,771 positions where safe paid

League 3

Random · a random legal play

Risk judgement 24.9 (95%: 22.3–27.4)

Risks found 24.8% · safe right 25.0%

Took 112 of 452 risks that paid · played safe in 121 of 484 positions where safe paid

Error rate 121.1 (95%: 113.3–128.5) · record 0–0–18

League 1

1,337 decisions

Model page →

Random (computed)

Risk judgement 27.0 (95%: 26.7–27.2)

Risks found 25.1% · safe right 28.8%

Took 3,300.3 of 13,140 risks that paid · played safe in 4,655.0 of 16,136 positions where safe paid

League 2

Random (computed)

Risk judgement 27.6 (95%: 27.1–28.0)

Risks found 25.1% · safe right 30.1%

Took 1,293.3 of 5,153 risks that paid · played safe in 2,036.5 of 6,771 positions where safe paid

League 3

How each cost is counted

API-equivalent (subscription). At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid.
Estimate: $47.092 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 8.2 h on 4 × NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h × 4 = $81.21; plus 1,556 answers from the hosted API at their recorded cost, $23.62; total $104.83 ÷ 2,226 decisions; left out: 38 gaps over 10 min between consecutive answers, 33.3 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
API price. At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid.
Estimate: $0.728 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.07 h on 4 × NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h × 4 = $0.65; plus 1,641 answers from the hosted API at their recorded cost, $0.61; total $1.26 ÷ 1,736 decisions; left out: 7 gaps over 10 min between consecutive answers, 30.8 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
API price (billed). At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid.
Estimate: $34.943 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.3 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.27 + 22.1 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $55.13 + 3.8 h on 2 × NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h × 2 = $18.75; total $74.15 ÷ 2,122 decisions; left out: 62 gaps over 10 min between consecutive answers, 134.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $8.215 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.34 + 6.0 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $14.83; total $15.17 ÷ 1,847 decisions; left out: 10 gaps over 10 min between consecutive answers, 151.2 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.999 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.8 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $2.02; total $2.02 ÷ 2,021 decisions; left out: 1 gaps over 10 min between consecutive answers, 0.7 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $1.285 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.1 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $2.76; total $2.76 ÷ 2,148 decisions; left out: 1 gaps over 10 min between consecutive answers, 5.3 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.794 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 6.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $1.38; total $1.38 ÷ 1,740 decisions; left out: 1 gaps over 10 min between consecutive answers, 32.5 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $39.023 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 5.0 h on NVIDIA DGX Spark (GB10) × $0.213/h = $1.07 + 32.6 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $81.19; plus 1 answers from the hosted API at their recorded cost, $0.00; total $82.26 ÷ 2,108 decisions; left out: 61 gaps over 10 min between consecutive answers, 122.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $4.551 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.06 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.01 + 3.7 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $9.11; total $9.13 ÷ 2,005 decisions; left out: 6 gaps over 10 min between consecutive answers, 117.9 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $1.545 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.2 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $3.09; total $3.09 ÷ 1,998 decisions; left out: 2 gaps over 10 min between consecutive answers, 4.9 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.293 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 2.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.54; total $0.54 ÷ 1,855 decisions; left out: 1 gaps over 10 min between consecutive answers, 14.6 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
list price; output not billed. At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid.
Estimate: $10.301 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 8.3 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $20.63; total $20.63 ÷ 2,003 decisions; left out: 26 gaps over 10 min between consecutive answers, 81.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
list price (billed); output not billed. At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid.
Estimate: $0.782 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.6 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $1.56; total $1.56 ÷ 1,993 decisions; no gap over 10 min. NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $12.227 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 10.0 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $24.80; total $24.80 ÷ 2,028 decisions; left out: 19 gaps over 10 min between consecutive answers, 131.1 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.685 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.6 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $1.40; total $1.40 ÷ 2,044 decisions; left out: 2 gaps over 10 min between consecutive answers, 1.0 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $5.736 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.34 + 4.4 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $11.01; total $11.35 ÷ 1,978 decisions; left out: 7 gaps over 10 min between consecutive answers, 243.2 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.072 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.7 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.14; total $0.14 ÷ 1,989 decisions; left out: 1 gaps over 10 min between consecutive answers, 12.5 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $11.168 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.6 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.12 + 9.0 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $22.39; total $22.51 ÷ 2,016 decisions; left out: 19 gaps over 10 min between consecutive answers, 252.8 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $6.593 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 2.3 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.49 + 5.2 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $13.04; total $13.53 ÷ 2,052 decisions; left out: 34 gaps over 10 min between consecutive answers, 255.1 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
API price (billed); output free. At each setup's price basis (a list price, or API-equivalent for a subscription); not a bill paid.
Estimate: $0.139 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.29; total $0.29 ÷ 2,108 decisions; left out: 1 gaps over 10 min between consecutive answers, 11.5 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.467 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $0.93; total $0.93 ÷ 1,980 decisions; left out: 1 gaps over 10 min between consecutive answers, 2.5 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $9.615 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 7.8 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $19.37; total $19.37 ÷ 2,015 decisions; left out: 22 gaps over 10 min between consecutive answers, 149.6 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.052 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.09; total $0.09 ÷ 1,771 decisions; left out: 1 gaps over 10 min between consecutive answers, 63.8 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.657 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $1.12; total $1.12 ÷ 1,710 decisions; no gap over 10 min. NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $38.716 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.1 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.02 + 34.1 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $84.88; total $84.90 ÷ 2,193 decisions; left out: 16 gaps over 10 min between consecutive answers, 97.1 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $9.132 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 7.2 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $17.83; total $17.83 ÷ 1,952 decisions; left out: 19 gaps over 10 min between consecutive answers, 87.9 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.049 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.10; total $0.10 ÷ 1,955 decisions; left out: 1 gaps over 10 min between consecutive answers, 34.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.054 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.12; total $0.12 ÷ 2,153 decisions; left out: 1 gaps over 10 min between consecutive answers, 64.1 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $12.267 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 9.9 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $24.61; total $24.61 ÷ 2,006 decisions; left out: 22 gaps over 10 min between consecutive answers, 147.3 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $7.185 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 5.8 h on NVIDIA RTX PRO 6000 Blackwell (96 GB) × $2.49/h = $14.38; total $14.38 ÷ 2,002 decisions; left out: 13 gaps over 10 min between consecutive answers, 257.6 h in all (each one listed in the data file). NVIDIA RTX PRO 6000 Blackwell (96 GB): $2.49/h per GPU. Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.058 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.10; total $0.10 ÷ 1,676 decisions; left out: 1 gaps over 10 min between consecutive answers, 30.9 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.049 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.08; total $0.08 ÷ 1,728 decisions; left out: 1 gaps over 10 min between consecutive answers, 55.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.054 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.5 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.10; total $0.10 ÷ 1,882 decisions; left out: 1 gaps over 10 min between consecutive answers, 54.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.052 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.09; total $0.09 ÷ 1,763 decisions; left out: 1 gaps over 10 min between consecutive answers, 53.0 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.051 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.09; total $0.09 ÷ 1,832 decisions; left out: 1 gaps over 10 min between consecutive answers, 31.4 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.057 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 0.4 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.08; total $0.08 ÷ 1,355 decisions; left out: 1 gaps over 10 min between consecutive answers, 31.7 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.
Estimate: $0.174 per 1,000 decisions (GPU time, at most). Counted as the wall clock from the first to the last answer, pauses over 10 min left out, × the hourly rate ÷ decisions. 1.2 h on NVIDIA DGX Spark (GB10) × $0.213/h = $0.25; total $0.25 ÷ 1,405 decisions; left out: 2 gaps over 10 min between consecutive answers, 42.1 h in all (each one listed in the data file). NVIDIA DGX Spark (GB10): CHF 4,000 ÷ (3 years × 8,760 h) = CHF 0.152/h purchase + CHF 0.2838/kWh (ewz.natur high rate) × 0.0791 kW (the chip's measured draw, not the whole box) = CHF 0.022/h electricity; CHF 0.175/h ≈ $0.213/h (CHF 1 = $1.2185, ECB 2026-11-10). Other setups' servers may have run on the same machine at the same time; each setup is charged the whole device, so this is an upper bound.

Learning: did they improve?

The leaderboard's Learning tab shows, for every model, its verdict where one is published, else its state in the learning lanes: running, waiting (and why), not tested (and the rule), or not in the learning plan.

How a model is judged. It answers the 200-position risk test 3 times with no memory (before = the mean of those empty runs), then plays 5 series of 10 games against GNUbg weak, writing its own lessons after each game, and takes the test again after game 10 of each series (after = the mean of the 5). Improved? is Yes when risks found rises by more than the improvement threshold (the spread, max − min, of its own 3 empty runs, at least 5 points) and right safety falls by no more than 5 points. This is an operational threshold, not a statistical test. Yes or No only once every run is complete.

Did they improve? Every result, test and lesson →

Same positions

Preview. The stored answers of the "Did they improve?" test without lessons (E1): 200 positions every model below answered; Muse Ember's answers come from its own run of the same test (12 Nov 2026, its league route and settings). The prompt and setting are the league row's (checked request by request against the league prompts). Engine's best play: the play loses under 0.001 against GNU Backgammon's best; an invalid answer counts as wrong. Tiers: an exact McNemar test on every pair, Holm-adjusted at 0.05; tier 1 is the model with the most best plays and every model not significantly different from it, then the same for the models left. Rank is the league rank.

Same positions: 15 models, 200 positions each, by tier
TierModelEngine's best play, of 200Invalid (counted wrong)League rank
1Claude Fable 5.1high10603
1Claude Opus 5.5high10605
1Claude Sonnet 5.5high10308
1Gemini 4 Argonhigh101452
1DeepSeek-V4.1-Flashlow · API + NVIDIA RTX PRO 6000 Blackwell · first release10016
1Gemini 3.8 Flashhigh9709
1Grok 4.7high9707
1Gemma 4 31Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark92013
1Muse Embermax88410
1Kimi K3.1 Prooff85011
1GLM-5.3low83018
2Qwen3.8-27Blow · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark78017
2Drex Max v1.167025
2Qwen3.8 Flashoff66023
2SubQ Snap 1.160034

Not asked yet: 26 ranked models (GPT-6 Astrahigh, GPT-6.1 Solhigh, Gemini Decision 1, d1-ultra, Mammoth-1, Drex MaxNVIDIA RTX PRO 6000 Blackwell, KnowLine-31B-Gen1.5NVIDIA DGX Spark (GB10), REAX 1.3, GLiDEAPI, Jev 1.13, Jev 1.14, GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell, Drex 2.0, Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell, celeris-1-decision, Elephant-1, Clef, Cygnet, Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark, SubQ Snap Preview, GPT-6 LunaDecisions API, Winnow-12B, Mercury Decide, LFM2.5-24B-A2Bon · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark, Perplexity Decider Flash v1.2, Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell). Asking them, one position per request: about $0.4 at list price on hosted APIs ($0.34 for those available now), 19.0 hours one after another, of which 17.7 hours on our GPUs. Not available now: GPT-6 Astrahigh (ChatGPT subscription quota until 13 Nov), GPT-6.1 Solhigh (ChatGPT subscription quota until 13 Nov), GPT-6 LunaDecisions API (no OpenAI API credit).

Models are too cautious.

56 of the 57 ranked models found a worthwhile risk (a risky play at least 0.001 better than the best safe play) in a smaller share of the risk-pays positions than the share of safety-pays positions they played safe: risks found runs from 18.7% to 53.3%, safe right from 16.3% to 97.6%. GNUbg weak, a weak setting of the engine (grey square), is close to even: 67.6% safe right, 66.6% risks found. Above the diagonal, a model succeeds more often when safety pays than when risk pays; the chart compares these two success rates, not how often a model plays risky at all.

0%0%25%25%50%50%75%75%100%100%Risks found: a worthwhile risk when risk paidSafe right: played safe when safe paidGPT-6 Astra: risks found 53.3%, safe right 95.5%Gemini 4 Argon: risks found 50.8%, safe right 97.4%Claude Fable 5.1: risks found 50.7%, safe right 94.0%GPT-6.1 Sol: risks found 50.6%, safe right 93.7%Claude Opus 5.5: risks found 43.4%, safe right 96.6%DeepSeek-V4.1-Flash: risks found 45.0%, safe right 93.0%Grok 4.7: risks found 44.0%, safe right 93.0%Claude Sonnet 5.5: risks found 41.8%, safe right 94.6%Gemini 3.8 Flash: risks found 36.1%, safe right 97.6%Muse Ember: risks found 37.6%, safe right 94.4%Kimi K3.1 Pro: risks found 42.4%, safe right 89.5%Gemini Decision 1: risks found 40.9%, safe right 88.5%Gemma 4 31B: risks found 31.9%, safe right 97.1%d1-ultra: risks found 42.0%, safe right 86.0%Mammoth-1: risks found 40.0%, safe right 86.4%Drex Max (NVIDIA RTX PRO 6000 Blackwell): risks found 39.3%, safe right 86.5%Sieve 31B v3: risks found 35.2%, safe right 90.2%Qwen3.8-27B: risks found 33.8%, safe right 90.2%GLM-5.3: risks found 30.7%, safe right 92.7%KnowLine-31B-Gen1.5 (NVIDIA DGX Spark (GB10)): risks found 28.8%, safe right 94.4%REAX 1.3: risks found 31.7%, safe right 89.4%GLiDE (API): risks found 36.9%, safe right 84.0%Turion Jev 31B: risks found 24.0%, safe right 95.1%Jev 1.13: risks found 27.0%, safe right 92.1%Qwen3.8 Flash: risks found 34.4%, safe right 84.6%Jev 1.14: risks found 29.1%, safe right 89.4%Drex Max v1.1: risks found 35.6%, safe right 82.5%GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell): risks found 33.0%, safe right 84.6%Drex 2.0: risks found 31.5%, safe right 84.8%Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell): risks found 33.4%, safe right 82.5%celeris-1-decision: risks found 23.9%, safe right 91.7%Elephant-1: risks found 28.4%, safe right 87.0%Clef: risks found 39.5%, safe right 75.6%StartLux-Decision-v1.1-4B-Instant: risks found 34.1%, safe right 80.4%Cygnet: risks found 27.3%, safe right 86.8%Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): risks found 33.1%, safe right 80.7%SubQ Snap 1.1: risks found 38.6%, safe right 74.8%SubQ Snap Preview: risks found 37.9%, safe right 75.1%Cerno 9B v1.2: risks found 39.2%, safe right 72.9%GPT-6 Luna (Decisions API): risks found 21.6%, safe right 89.7%d1-2B (NVIDIA RTX PRO 6000 Blackwell): risks found 33.6%, safe right 76.8%Winnow-12B: risks found 33.7%, safe right 76.6%StartLux-Decision-v1.1-2B-Instant: risks found 33.7%, safe right 74.0%Mercury Decide: risks found 18.7%, safe right 88.1%d1-Omni-2B (NVIDIA RTX PRO 6000 Blackwell): risks found 30.4%, safe right 76.0%LFM2.5-24B-A2B: risks found 31.5%, safe right 73.4%KnowLine-2B-Gen1: risks found 30.0%, safe right 72.3%KnowLine 4B Gen5.5: risks found 34.1%, safe right 65.4%Perplexity Decider Flash v1.2: risks found 31.3%, safe right 66.9%Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell): risks found 37.3%, safe right 59.8%GLiDE 2B: risks found 33.0%, safe right 60.3%KnowLine-0.8B-Gen1: risks found 30.8%, safe right 59.6%StartLux-Decision-v1.1-0.8B-Instant: risks found 38.3%, safe right 50.9%OneJev1.6-0.8B: risks found 27.6%, safe right 57.4%GLiDE 0.8B: risks found 30.4%, safe right 53.1%GLiDE 0.4B: risks found 18.8%, safe right 39.0%GLiClass v4 Edge: risks found 26.8%, safe right 16.3%GNUbg weak: risks found 66.6%, safe right 67.6%GNUbg weak
Every model's two values

Too safe

Each model's row of the table: the share of its right-safety calls and of the risks it found, as in the table

  1. GPT-6 Astra · late entry95.5% | 53.3%
  2. Gemini 4 Argon · League 397.4% | 50.8%
  3. Claude Fable 5.1 · League 394.0% | 50.7%
  4. GPT-6.1 Sol · League 393.7% | 50.6%
  5. Claude Opus 5.5 · League 396.6% | 43.4%
  6. DeepSeek-V4.1-Flash · League 293.0% | 45.0%
  7. Grok 4.7 · League 393.0% | 44.0%
  8. Claude Sonnet 5.5 · League 294.6% | 41.8%
  9. Gemini 3.8 Flash · League 297.6% | 36.1%
  10. Muse Ember · late entry94.4% | 37.6%
  11. Kimi K3.1 Pro · League 389.5% | 42.4%
  12. Gemini Decision 1 · late entry88.5% | 40.9%
  13. Gemma 4 31B · League 297.1% | 31.9%
  14. d1-ultra · League 286.0% | 42.0%
  15. Mammoth-1 · late entry86.4% | 40.0%
  16. Drex MaxNVIDIA RTX PRO 6000 Blackwell · late entry86.5% | 39.3%
  17. Sieve 31B v3 · late entry90.2% | 35.2%
  18. Qwen3.8-27B · League 290.2% | 33.8%
  19. GLM-5.3 · League 292.7% | 30.7%
  20. KnowLine-31B-Gen1.5NVIDIA DGX Spark (GB10) · late entry94.4% | 28.8%
  21. REAX 1.3 · late entry89.4% | 31.7%
  22. GLiDEAPI · League 284.0% | 36.9%
  23. Turion Jev 31B · late entry95.1% | 24.0%
  24. Jev 1.13 · League 192.1% | 27.0%
  25. Qwen3.8 Flash · League 284.6% | 34.4%
  26. Jev 1.14 · League 289.4% | 29.1%
  27. Drex Max v1.1 · late entry82.5% | 35.6%
  28. GLiDEv1.2 · NVIDIA RTX PRO 6000 Blackwell · League 184.6% | 33.0%
  29. Drex 2.0 · late entry84.8% | 31.5%
  30. Perplexity Decider 27B v1.2NVIDIA RTX PRO 6000 Blackwell · League 182.5% | 33.4%
  31. celeris-1-decision · League 191.7% | 23.9%
  32. Elephant-1 · late entry87.0% | 28.4%
  33. Clef · League 275.6% | 39.5%
  34. StartLux-Decision-v1.1-4B-Instant · late entry80.4% | 34.1%
  35. Cygnet · League 186.8% | 27.3%
  36. Perplexity Decider 27B v1.1NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark · League 180.7% | 33.1%
  37. SubQ Snap 1.1 · late entry74.8% | 38.6%
  38. SubQ Snap Preview · League 175.1% | 37.9%
  39. Cerno 9B v1.2 · late entry72.9% | 39.2%
  40. GPT-6 LunaDecisions API · League 189.7% | 21.6%
  41. d1-2BNVIDIA RTX PRO 6000 Blackwell · late entry76.8% | 33.6%
  42. Winnow-12B · League 176.6% | 33.7%
  43. StartLux-Decision-v1.1-2B-Instant · late entry74.0% | 33.7%
  44. Mercury Decide · League 188.1% | 18.7%
  45. d1-Omni-2BNVIDIA RTX PRO 6000 Blackwell · late entry76.0% | 30.4%
  46. LFM2.5-24B-A2B · League 173.4% | 31.5%
  47. KnowLine-2B-Gen1 · late entry72.3% | 30.0%
  48. KnowLine 4B Gen5.5 · late entry65.4% | 34.1%
  49. Perplexity Decider Flash v1.2 · League 166.9% | 31.3%
  50. Open-Jev-27B-v1.1NVIDIA RTX PRO 6000 Blackwell · League 159.8% | 37.3%
  51. GLiDE 2B · late entry60.3% | 33.0%
  52. KnowLine-0.8B-Gen1 · late entry59.6% | 30.8%
  53. StartLux-Decision-v1.1-0.8B-Instant · late entry50.9% | 38.3%
  54. OneJev1.6-0.8B · late entry57.4% | 27.6%
  55. GLiDE 0.8B · late entry53.1% | 30.4%
  56. GLiDE 0.4B · late entry39.0% | 18.8%
  57. GLiClass v4 Edge · late entry16.3% | 26.8%
  58. GNUbg weak · League 167.6% | 66.6%

right safety: played safe when safe was right risks found: took the risk when it paid GNUbg weak, a program: the grey reference pair Bars in the maker's colour, the same rows as the table: OpenAIGoogleAnthropicxAIDeepSeekMoonshotLiquid AIBase LabsAlibabaInceptionNace.AICloudflareZ.aiFastinoPerplexitySubquadraticCelerisEldanRingREAXTypeSafeOmniJevZefan CaiBlockbrain LabsPelaAISthānika AITurion.AIApertix AIStartLuxMetaKnowledgator

OpenAIGoogleAnthropicxAIDeepSeekMoonshotLiquid AIBase LabsAlibabaInceptionNace.AICloudflareZ.aiFastinoPerplexitySubquadraticCelerisEldanRingREAXTypeSafeOmniJevZefan CaiBlockbrain LabsPelaAISthānika AITurion.AIApertix AIStartLuxMetaKnowledgator GNUbg weak (program, reference) Dashed diagonal: as often right on both

Not paired: models met different positions in different leagues. How risks found and safe right are counted

Risk judgement by release date

A release month is documented for 24 of the 57 ranked models; only those are shown, each at the month of its maker's own dated statement. Makers with models in different months, earliest first: Google 64.5, 66.9, 74.1; Fastino 60.4, 58.8. That is not a consistent increase across the makers shown.

45505560657075AprMayJunJulAugSepOct2026Risk judgementGemma 4 31B (Google), released Apr 2026: risk judgement 64.5Gemini 3.8 Flash (Google), released Sep 2026: risk judgement 66.9Gemini 4 Argon (Google), released Oct 2026: risk judgement 74.1GoogleGLM-5.3 (Z.ai), released Aug 2026: risk judgement 61.7Qwen3.8-27B (Alibaba), released Aug 2026: risk judgement 62.0Claude Fable 5.1 (Anthropic), released Sep 2026: risk judgement 72.3Claude Opus 5.5 (Anthropic), released Sep 2026: risk judgement 70.0Claude Sonnet 5.5 (Anthropic), released Sep 2026: risk judgement 68.2DeepSeek-V4.1-Flash (DeepSeek), released Sep 2026: risk judgement 69.0GLiDE (API) (Fastino), released Sep 2026: risk judgement 60.4GLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell) (Fastino), released Oct 2026: risk judgement 58.8FastinoGPT-6 Astra (OpenAI), released Sep 2026: risk judgement 74.4GPT-6 Luna (Decisions API) (OpenAI), released Sep 2026: risk judgement 55.7GPT-6.1 Sol (OpenAI), released Sep 2026: risk judgement 72.2Grok 4.7 (xAI), released Sep 2026: risk judgement 68.5celeris-1-decision (Celeris), released Oct 2026: risk judgement 57.8Clef (Cloudflare), released Oct 2026: risk judgement 57.5Jev 1.14 (TypeSafe), released Oct 2026: risk judgement 59.3Kimi K3.1 Pro (Moonshot), released Oct 2026: risk judgement 65.9LFM2.5-24B-A2B (Liquid AI), released Oct 2026: risk judgement 52.4d1-ultra (Liquid AI), released Oct 2026: risk judgement 64.0Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark) (Perplexity), released Oct 2026: risk judgement 56.9Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell) (Perplexity), released Oct 2026: risk judgement 57.9Perplexity Decider Flash v1.2 (Perplexity), released Oct 2026: risk judgement 49.1

OpenAIGoogleAnthropicxAIDeepSeekMoonshotLiquid AIAlibabaCloudflareZ.aiFastinoPerplexityCelerisTypeSafe

Models without a documented release month are left out. How release months are dated

With the numbers on screen, every model asked played safer

Asked which of two plays is better on the same rollout-confirmed positions, all 8 models chose the safe play more often when the prompt showed each play's computed numbers than from the board alone: the bias (missed minus needless) moves from -22 pp to +92 pp for Jev 1.14 and from +46 pp to +56 pp for Claude Opus 5.5high. From the board alone, 4 of them (celeris-1-decision, Jev 1.14, d1-ultra, Kimi K3.1 Prooff) took more needless risks than they missed.

  1. Jev 1.14+92 pp | -22 pp
  2. celeris-1-decision+91 pp | -14 pp
  3. GLM-5.3low+78 pp | +18 pp
  4. d1-ultra+74 pp | -21 pp
  5. Kimi K3.1 Prooff+70 pp | -25 pp
  6. Claude Sonnet 5.5high+66 pp | +42 pp
  7. Claude Opus 5.5high+56 pp | +46 pp
  8. GPT-6.1 Solhigh+54 pp | +4 pp

with numbers (the league prompt, each play's computed consequences) board only (notation, no computed number) Bias: missed minus needless, Q1 of Clear risk Part 2

Preliminary: two plays shown, not league play. Clear risk (preliminary)

What thinking adds

9 models played two or more reasoning settings in the same league. At the highest setting, 9 of the 9 had a lower error rate than at the lowest, and 7 of the 9 a higher risk judgement. GPT-6 Astra, low to high: risk judgement 75.8, 74.7, 74.4; error rate 11.8, 11.1, 8.8.

Risk judgement (higher is better)Error rate, PR (lower is better)40506070800204060LFM2.5-24B-A2B · L1LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): 45.7LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 52.4LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell): 61.5LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 47.7Claude Sonnet 5.5 · L2Claude Sonnet 5.5 (low): 63.5Claude Sonnet 5.5 (high): 68.2Claude Sonnet 5.5 (low): 20.1Claude Sonnet 5.5 (high): 16.7DeepSeek-V4.1-Flash (1017) · L2DeepSeek-V4.1-Flash (off · API · 1017): 58.5DeepSeek-V4.1-Flash (low · API · 1017): 68.5DeepSeek-V4.1-Flash (off · API · 1017): 35.2DeepSeek-V4.1-Flash (low · API · 1017): 17.1Gemini 3.8 Flash · L2Gemini 3.8 Flash (low): 65.1Gemini 3.8 Flash (high): 66.9Gemini 3.8 Flash (low): 16.3Gemini 3.8 Flash (high): 14.9Gemma 4 31B · L2Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 62.6Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 64.5Gemma 4 31B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 22.5Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark): 19.6Claude Opus 5.5 · L3Claude Opus 5.5 (low): 66.9Claude Opus 5.5 (high): 70.0Claude Opus 5.5 (low): 15.4Claude Opus 5.5 (high): 14.4GPT-6 Astra · L3GPT-6 Astra (low): 75.8GPT-6 Astra (medium): 74.7GPT-6 Astra (high): 74.4GPT-6 Astra (low): 11.8GPT-6 Astra (medium): 11.1GPT-6 Astra (high): 8.8GPT-6.1 Sol · L3GPT-6.1 Sol (low): 72.7GPT-6.1 Sol (high): 72.2GPT-6.1 Sol (low): 12.1GPT-6.1 Sol (high): 9.9Gemini 4 Argon · L3Gemini 4 Argon (low): 69.2Gemini 4 Argon (high): 74.1Gemini 4 Argon (low): 16.5Gemini 4 Argon (high): 12.6

One row per model and league; hollow dots: its lower settings, filled dot: its highest setting. OpenAIGoogleAnthropicDeepSeekLiquid AI

Not paired: each setting played its own games; these are observed differences. Every setting, league by league

A new starting position

Each model was asked Nackgammon positions and standard positions matched to them. Compared group against group (one group's drop minus another's; a group is the setups from one original league: League 1, 2, 3), every 99% interval includes 0, so the test does not show the groups' drops to differ. Random play's own drop is −4.6 PR on the main 200 pairs: the Nackgammon positions were slightly easier for random play. Outside the groups, the late entry Elephant-1 dropped (+8.66 PR, 97.5% interval +0.42 to +16.02). Unclear means the test can’t tell, not that the scores are equal.

00101020203030404050506060707080809090100100Standard PR, main 200 pairsNackgammon PR, main 200 pairsGLiDE (v1.2 · NVIDIA RTX PRO 6000 Blackwell) (Open · 27B · League 1 group): standard PR 22.74, Nackgammon PR 24.91GLiDE (API) (Closed · Efficient · League 1 group): standard PR 22.93, Nackgammon PR 23.38d1-ultra (Closed · Efficient · League 1 group): standard PR 24.25, Nackgammon PR 28.27Jev 1.13 (Closed · Efficient · League 1 group): standard PR 26.53, Nackgammon PR 24.17GPT-6 Luna (Decisions API) (Closed · Efficient · League 1 group): standard PR 28.27, Nackgammon PR 34.11celeris-1-decision (Closed · Efficient · League 1 group): standard PR 29.71, Nackgammon PR 28.47Perplexity Decider 27B v1.1 (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark) (Open · 27B · League 1 group): standard PR 30.42, Nackgammon PR 28.64Cygnet (NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark) (Open · 12B · League 1 group): standard PR 31.33, Nackgammon PR 24.48SubQ Snap Preview (Closed · Efficient · League 1 group): standard PR 31.91, Nackgammon PR 29.66Mercury Decide (Closed · Efficient · League 1 group): standard PR 32.74, Nackgammon PR 28.46Winnow-12B (NVIDIA RTX PRO 6000 Blackwell) (Open · 12B · League 1 group): standard PR 34.33, Nackgammon PR 33.43Perplexity Decider 27B v1.2 (NVIDIA RTX PRO 6000 Blackwell) (Open · 27B · League 1 group): standard PR 34.68, Nackgammon PR 22.80Perplexity Decider Flash v1.2 (NVIDIA RTX PRO 6000 Blackwell) (Open · 4.0B · League 1 group): standard PR 39.77, Nackgammon PR 42.86Open-Jev-27B-v1.1 (NVIDIA RTX PRO 6000 Blackwell) (Open · 27B · League 1 group): standard PR 52.08, Nackgammon PR 40.96GPT-6.1 Sol (high) (Closed · Workhorse · League 2 group): standard PR 9.88, Nackgammon PR 8.43GPT-6.1 Sol (low) (Closed · Workhorse · League 2 group): standard PR 11.15, Nackgammon PR 9.90Claude Sonnet 5.5 (low) (Closed · Workhorse · League 2 group): standard PR 13.96, Nackgammon PR 14.50Gemini 3.8 Flash (low) (Closed · Workhorse · League 2 group): standard PR 15.92, Nackgammon PR 11.81Gemma 4 31B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark) (Open · 31B · League 2 group): standard PR 18.35, Nackgammon PR 13.38DeepSeek-V4.1-Flash (off · API + NVIDIA RTX PRO 6000 Blackwell · first release) (Open · 552B MoE · 16B active · League 2 group): standard PR 23.91, Nackgammon PR 27.90Qwen3.8-27B (off · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark) (Open · 27B · League 2 group): standard PR 28.13, Nackgammon PR 28.43LFM2.5-24B-A2B (on · NVIDIA RTX PRO 6000 Blackwell + NVIDIA DGX Spark) (Open · 24B MoE · 2.5B active · League 2 group): standard PR 38.20, Nackgammon PR 37.55LFM2.5-24B-A2B (off · NVIDIA RTX PRO 6000 Blackwell) (Open · 24B MoE · 2.5B active · League 2 group): standard PR 50.75, Nackgammon PR 53.98Claude Fable 5.1 (high) (Closed · Flagship · League 3 group): standard PR 9.13, Nackgammon PR 9.51GPT-6 Astra (medium) (Closed · Flagship · League 3 group): standard PR 9.66, Nackgammon PR 7.30Grok 4.7 (high) (Closed · Workhorse · League 3 group): standard PR 15.20, Nackgammon PR 14.53GPT-6.1 Terra (no group (exploratory)): standard PR 13.52, Nackgammon PR 14.40Gemini 4 Argon (low) (Closed · Flagship · no group (exploratory)): standard PR 13.79, Nackgammon PR 15.71GNUbg weak (yardstick): standard PR 27.69, Nackgammon PR 24.13GNUbg weakRandom (yardstick): standard PR 87.45, Nackgammon PR 82.89Random

OpenAIGoogleAnthropicxAIDeepSeekLiquid AIAlibabaInceptionFastinoPerplexitySubquadraticCelerisEldanRingTypeSafeZefan CaiBlockbrain Labs GNUbg weak, Random (programs) Above the dashed diagonal: higher error on Nackgammon.

Paired positions, not league play; a drop mixes the new start with how hard the positions are. The Nackgammon test

Does the model learn from its games?

Status: Done. No arm reached the pre-registered signal threshold (plan section 8).

Gemma 4 31B: risks found, % (higher is better)25%30%35%Before game 1After game 5After game 10no memory, Before game 1: 25.0%full records, Before game 1: 25.0%full records, After game 5: 28.8%full records, After game 10: 28.3%own lessons, Before game 1: 25.0%own lessons, After game 5: 34.6%own lessons, After game 10: 33.3%budget (empty: no game fit), Before game 1: 25.0%budget (empty: no game fit), After game 5: 28.7%budget (empty: no game fit), After game 10: 27.8%

no memoryfull recordsown lessonsbudget (empty: no game fit)

Swipe for more columns →

Risks found (%) on the 200 evaluation positions: means over each arm's finished series; before game 1 is one shared run
ArmSeriesBefore game 1After game 5After game 10
no memory1025.0T0 onlyT0 only
full records1025.028.828.3
own lessons1025.034.633.3
budget (empty: no game fit)1025.028.727.8

no memory is tested before game 1 only (the shared run): its line is drawn flat and dashed at that value.

One model, a small pilot; the plan, every number and the lessons are on its page. Learning from experience

How the scores work

  1. Models play full games of backgammon. GNU Backgammon scores every move.
  2. Risk judgement: did the model find a worthwhile risk when risk paid, and play safe when safe was right? 50 = always plays safe, 100 = every call right.
  3. Error rate: how much a move gives away, judged by the engine (lower is better).
  4. Everything is public: every position, every legal play, the engine's value for each.
  5. Leagues played different positions, so close scores across leagues are not a clear ranking.

More in About