Decision model
Clef-flash
1 question per request
45.7% of the 2,198 questions right (43.6% to 47.8%, 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-flash- Parameters
- 9.5B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Cloudflare's 9B decision model, post-trained from Qwen3.5-9B
- Confidence from
- Option score
- Run
- 139, 7 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 | 45.7% | 43.6% to 47.8% | 1,005 of 2,198 |
| Reading | 73.0% | 69.5% to 76.3% | 479 of 656 |
| Grammar | 34.1% | 31.8% to 36.5% | 526 of 1,542 |
Comparison group: E. 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
- 15.5 points
- AUROC, all questions
- 0.74 0.72 to 0.76
- AUROC, reading
- 0.77 0.73 to 0.81
- AUROC, grammar
- 0.64 0.61 to 0.67
- Brier score
- 0.230
- Accuracy of its surest answers
- most confident fifth: 78.6% right; most confident half: 63.6% right; most confident four fifths: 51.8% right
- Typical score of a right answer
- 76.1%
- Typical score of a wrong answer
- 49.1%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.075 s (this run)
- Latency, 90th percentile
- 0.081 s
- Requests at once
- 1
- Throughput
- 795.3 questions per minute of request time
- Request time
- 2.8 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
- $0.017
- Whole run
- $0.038
- Basis
- Estimated at Cloudflare's list price for the same model; our hardware cost is not measured.
- Price
- $0.038 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 | 0.9% | −0.08 points | 0%§ |
| GPT-6 Astrahigh reasoning · 16 questions per request | 0.9% | −0.08 points | 1%§ |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 0.9% | −0.08 points | -0%§ |
| GPT-6 Astralow reasoning · 16 questions per request | 0.9% | −0.08 points | 1%§ |
| GPT-6 Astralow reasoning · 1 question per request | 1.5% | −0.09 points | 1%§ |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 0.9% | −0.08 points | -1%§ |
| GPT-6.1 Sollow reasoning · 1 question per request | 1.6% | −0.10 points | 1%§ |
| GPT-6.1 Sollow reasoning · 16 questions per request | 1.5% | −0.09 points | -0%§ |
| Claude Opus 5.5high reasoning · 16 questions per request | 1.5% | −0.09 points | 1%§ |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 1.5% | −0.09 points | -2%§ |
| Claude Opus 5.5low reasoning · 1 question per request | 1.4% | −0.09 points | 1%§ |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 0.9% | −0.08 points | -6%§ |
| GPT-6.1 Terralow reasoning · 16 questions per request | 1.5% | −0.09 points | -3%§ |
| Claude Opus 5.5low reasoning · 16 questions per request | 1.6% | −0.10 points | 1%§ |
| Claude Sonnet 5.5high reasoning · 1 question per request | 1.6% | −0.10 points | 1%§ |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 1.9% | −0.13 points | 1%§ |
§ Counts the first model at its maker’s list price for the same model; our hardware cost is not measured.