Decision model
Metask-Jev 4B
1 question per request
40.8% of the 2,198 questions right (38.7% to 42.8%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- Metask-Jev server, self-hosted
- Paid as
- Our own hardware
- Model ID
metask-jev-4b-policy-mix- Parameters
- 4.5B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Fine-tune of Qwen3.5-4B (merged LoRA)
- Confidence from
- Option score
- Run
- 126, 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 | 40.8% | 38.7% to 42.8% | 896 of 2,198 |
| Reading | 62.8% | 59.0% to 66.4% | 412 of 656 |
| Grammar | 31.4% | 29.1% to 33.7% | 484 of 1,542 |
Comparison group: F. 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
- 5.8 points
- AUROC, all questions
- 0.73 0.70 to 0.75
- AUROC, reading
- 0.80 0.77 to 0.84
- AUROC, grammar
- 0.62 0.58 to 0.65
- Brier score
- 0.198
- Accuracy of its surest answers
- most confident fifth: 81.3% right; most confident half: 55.8% right; most confident four fifths: 45.0% right
- Typical score of a right answer
- 55.2%
- Typical score of a wrong answer
- 33.2%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.090 s (this run)
- Latency, 90th percentile
- 0.098 s
- Requests at once
- 1
- Throughput
- 660.6 questions per minute of request time
- Request time
- 3.3 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
- 518
- 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 | 3.2% | −0.12 points | 3%† |
| GPT-6 Astrahigh reasoning · 16 questions per request | 2.9% | −0.09 points | 3%† |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 2.9% | −0.09 points | 3%† |
| GPT-6 Astralow reasoning · 16 questions per request | 2.9% | −0.09 points | 3%† |
| GPT-6 Astralow reasoning · 1 question per request | 2.9% | −0.09 points | 3%† |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 5.0% | −0.17 points | 5%† |
| GPT-6.1 Sollow reasoning · 1 question per request | 3.2% | −0.12 points | 3%† |
| GPT-6.1 Sollow reasoning · 16 questions per request | 2.9% | −0.09 points | 3%† |
| Claude Opus 5.5high reasoning · 16 questions per request | 3.5% | −0.12 points | 4%† |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 3.2% | −0.12 points | 3%† |
| Claude Opus 5.5low reasoning · 1 question per request | 3.0% | −0.09 points | 3%† |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 3.2% | −0.12 points | 3%† |
| GPT-6.1 Terralow reasoning · 16 questions per request | 4.8% | −0.15 points | 5%† |
| Claude Opus 5.5low reasoning · 16 questions per request | 3.5% | −0.12 points | 4%† |
| Claude Sonnet 5.5high reasoning · 1 question per request | 3.5% | −0.12 points | 4%† |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 7.2% | −0.19 points | 7%† |
† Leaves out the first model’s hardware cost, which we have not estimated.