Decision model
Cygnet
1 question per request
55.7% of the 2,198 questions right (53.6% to 57.8%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- Cygnet decision server on vLLM, self-hosted
- Paid as
- Our own hardware
- Model ID
cygnet- Parameters
- 12B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Gemma 4 12B IT, unchanged, read out by option letter
- Confidence from
- Option score
- Run
- 127, 6 October 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
- Harness seed
- 0
- Harness commit
- not recorded
- Prompt version
- not recorded
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 55.7% | 53.6% to 57.8% | 1,224 of 2,198 |
| Reading | 82.8% | 79.7% to 85.5% | 543 of 656 |
| Grammar | 44.2% | 41.7% to 46.7% | 681 of 1,542 |
Comparison group: zABC. Compare with another setup.
Accuracy by topic
Confidence
- Answers kept at 99% or more
- 2, of which 0 wrong (share right: 34.2% to 100.0%, 95% interval); 2,196 skipped, the model less than 99% sure
- Calibration gap
- 13.8 points
- AUROC, all questions
- 0.80 0.79 to 0.82
- AUROC, reading
- 0.88 0.86 to 0.91
- AUROC, grammar
- 0.72 0.69 to 0.74
- Brier score
- 0.203
- Accuracy of its surest answers
- most confident fifth: 97.3% right; most confident half: 78.4% right; most confident four fifths: 63.2% right
- Typical score of a right answer
- 83.2%
- Typical score of a wrong answer
- 57.8%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.036 s (this run)
- Latency, 90th percentile
- 0.045 s
- Requests at once
- 1
- Throughput
- 1619.6 questions per minute of request time
- Request time
- 1.4 minutes, over the requests whose answers were kept
Latency: the time from sending one request to receiving its answer, measured from our computer, network and any retries within the request included. Request time: the time at least one request was in flight, over the requests whose answers were kept. It is not the run's elapsed time: it leaves out requests that failed and any pause before a run was resumed.
Cost
- $ per 1,000 questions
- Not measured
- Basis
- No dollar estimate; see hardware and time
- Input tokens per question
- 391
- Output tokens per question
- 1
Going first in front of an LLM
Replay of recorded answers, cutoff chosen on half the questions, averaged over the splits. Savings are estimates from API prices; the Method page explains how.
| LLM behind it | Share taken | Accuracy change | Saving |
|---|---|---|---|
| Gemini 3.8 Flashhigh reasoning · 16 questions per request | 11.7% | −0.08 points | 12%† |
| GPT-6 Astrahigh reasoning · 16 questions per request | 11.7% | −0.08 points | 12%† |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 11.7% | −0.08 points | 12%† |
| GPT-6 Astralow reasoning · 16 questions per request | 11.7% | −0.08 points | 12%† |
| GPT-6 Astralow reasoning · 1 question per request | 11.8% | −0.08 points | 12%† |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 11.7% | −0.09 points | 12%† |
| GPT-6.1 Sollow reasoning · 1 question per request | 13.7% | −0.10 points | 14%† |
| GPT-6.1 Sollow reasoning · 16 questions per request | 11.8% | −0.08 points | 12%† |
| Claude Opus 5.5high reasoning · 16 questions per request | 11.9% | −0.08 points | 12%† |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 11.9% | −0.08 points | 12%† |
| Claude Opus 5.5low reasoning · 1 question per request | 11.9% | −0.08 points | 12%† |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 11.9% | −0.09 points | 12%† |
| GPT-6.1 Terralow reasoning · 16 questions per request | 14.0% | −0.11 points | 14%† |
| Claude Opus 5.5low reasoning · 16 questions per request | 12.7% | −0.10 points | 13%† |
| Claude Sonnet 5.5high reasoning · 1 question per request | 13.8% | −0.11 points | 14%† |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 24.2% | −0.16 points | 24%† |
† Leaves out the first model’s hardware cost, which we have not estimated.