Decision model
Winnow-12B
1 question per request
55.5% of the 2,198 questions right (53.4% to 57.5%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- Winnow server (llama.cpp), self-hosted
- Paid as
- Our own hardware
- Model ID
winnow-12b- Parameters
- 11.9B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); Q8_0
- Note
- Fine-tune of Gemma 4 12B IT (Q8_0 GGUF)
- Confidence from
- Option score
- Run
- 128, 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 | 55.5% | 53.4% to 57.5% | 1,219 of 2,198 |
| Reading | 82.0% | 78.9% to 84.8% | 538 of 656 |
| Grammar | 44.2% | 41.7% to 46.7% | 681 of 1,542 |
Comparison group: ABC. Compare with another setup.
Accuracy by topic
Confidence
- Answers kept at 99% or more
- 163, of which 1 wrong (share right: 96.6% to 99.9%, 95% interval); 2,035 skipped, the model less than 99% sure
- Calibration gap
- 19.7 points
- AUROC, all questions
- 0.79 0.77 to 0.81
- AUROC, reading
- 0.85 0.82 to 0.89
- AUROC, grammar
- 0.70 0.67 to 0.73
- Brier score
- 0.237
- Accuracy of its surest answers
- most confident fifth: 97.0% right; most confident half: 76.9% right; most confident four fifths: 62.1% right
- Typical score of a right answer
- 94.5%
- Typical score of a wrong answer
- 63.5%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.059 s (this run)
- Latency, 90th percentile
- 0.070 s
- Requests at once
- 1
- Throughput
- 1002.4 questions per minute of request time
- Request time
- 2.2 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
- 404
- Output tokens per question
- 0
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 | 10.6% | −0.08 points | 11%† |
| GPT-6 Astrahigh reasoning · 16 questions per request | 10.6% | −0.08 points | 11%† |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 9.2% | −0.09 points | 9%† |
| GPT-6 Astralow reasoning · 16 questions per request | 9.2% | −0.09 points | 9%† |
| GPT-6 Astralow reasoning · 1 question per request | 9.2% | −0.09 points | 9%† |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 13.4% | −0.11 points | 13%† |
| GPT-6.1 Sollow reasoning · 1 question per request | 9.3% | −0.09 points | 9%† |
| GPT-6.1 Sollow reasoning · 16 questions per request | 10.7% | −0.09 points | 11%† |
| Claude Opus 5.5high reasoning · 16 questions per request | 10.6% | −0.08 points | 11%† |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 10.6% | −0.08 points | 11%† |
| Claude Opus 5.5low reasoning · 1 question per request | 9.3% | −0.09 points | 9%† |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 10.6% | −0.08 points | 11%† |
| GPT-6.1 Terralow reasoning · 16 questions per request | 11.8% | −0.12 points | 12%† |
| Claude Opus 5.5low reasoning · 16 questions per request | 15.6% | −0.10 points | 16%† |
| Claude Sonnet 5.5high reasoning · 1 question per request | 10.3% | −0.12 points | 10%† |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 23.2% | −0.13 points | 23%† |
† Leaves out the first model’s hardware cost, which we have not estimated.