Decision model
Clef
1 question per request
57.3% of the 2,198 questions right (55.2% to 59.3%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- Clef server, self-hosted
- Paid as
- Our own hardware
- Model ID
clef- Parameters
- 27.5B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Cloudflare's decision model, post-trained from Qwen3.8-27B
- Confidence from
- Option score
- Run
- 125, 5 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 | 57.3% | 55.2% to 59.3% | 1,259 of 2,198 |
| Reading | 85.1% | 82.1% to 87.6% | 558 of 656 |
| Grammar | 45.5% | 43.0% to 48.0% | 701 of 1,542 |
Comparison group: xyzAB. Compare with another setup.
Accuracy by topic
Confidence
- Answers kept at 99% or more
- 0, of which 0 wrong; 2,198 skipped, the model less than 99% sure
- Calibration gap
- 10.7 points
- AUROC, all questions
- 0.79 0.77 to 0.80
- AUROC, reading
- 0.88 0.85 to 0.90
- AUROC, grammar
- 0.69 0.66 to 0.72
- Brier score
- 0.199
- Accuracy of its surest answers
- most confident fifth: 97.3% right; most confident half: 78.2% right; most confident four fifths: 64.3% right
- Typical score of a right answer
- 85.7%
- Typical score of a wrong answer
- 52.1%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.205 s (this run)
- Latency, 90th percentile
- 0.248 s
- Requests at once
- 1
- Throughput
- 285.0 questions per minute of request time
- Request time
- 7.7 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
- $0.11
- Whole run
- $0.24
- Basis
- Estimated at Cloudflare's list price for the same model; our hardware cost is not measured.
- Price
- $0.24 input, output free per 1M tokens (source, 2026-10-18)
- Input tokens per question
- 457
- 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 | 9.5% | −0.11 points | 5%§ |
| GPT-6 Astrahigh reasoning · 16 questions per request | 11.2% | −0.14 points | 9%§ |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 11.3% | −0.15 points | 5%§ |
| GPT-6 Astralow reasoning · 16 questions per request | 11.2% | −0.14 points | 9%§ |
| GPT-6 Astralow reasoning · 1 question per request | 9.5% | −0.11 points | 9%§ |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 13.1% | −0.12 points | 2%§ |
| GPT-6.1 Sollow reasoning · 1 question per request | 14.3% | −0.15 points | 12%§ |
| GPT-6.1 Sollow reasoning · 16 questions per request | 11.2% | −0.14 points | 1%§ |
| Claude Opus 5.5high reasoning · 16 questions per request | 9.5% | −0.11 points | 8%§ |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 9.5% | −0.11 points | -12%§ |
| Claude Opus 5.5low reasoning · 1 question per request | 10.9% | −0.14 points | 10%§ |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 10.8% | −0.14 points | -30%§ |
| GPT-6.1 Terralow reasoning · 16 questions per request | 9.8% | −0.12 points | -16%§ |
| Claude Opus 5.5low reasoning · 16 questions per request | 12.0% | −0.14 points | 10%§ |
| Claude Sonnet 5.5high reasoning · 1 question per request | 13.6% | −0.17 points | 12%§ |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 22.8% | −0.19 points | 20%§ |
§ Counts the first model at its maker’s list price for the same model; our hardware cost is not measured.