Decision model
Drex Max v1
1 question per request
67.7% of the 2,198 questions right (65.8% to 69.7%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- Drex Max server (Nace's loader, drex_max), self-hosted, BF16
- Paid as
- Our own hardware
- Model ID
drex-max-v1- Parameters
- 26.1B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Nace.AI's released Drex Max weights (nace-ai/drex-max-v1 at cac65e1), the original Drex Max, a fine-tune of Qwen3.8-27B that also reads images. It ran on our hardware through Nace's own loader (drex_max, from nace-ai/drex-decision-models at e206b27), which makes 16 of the 64 layers attend non-causally, behind our HTTP server (servers/drex_max_server.py). One request at a time, as for the other setups.
- Confidence from
- Option score
- Run
- 322, 31 October 2026
Settings
- Reasoning
- off
- Request format
- 1 question per request
- Requests at once
- 1
- Temperature
- not recorded
- Harness seed
- 0
- Harness commit
61f3824138e9- Prompt version
254597b8be16
Results
| Questions | Accuracy | 95% interval | Right |
|---|---|---|---|
| All questions | 67.7% | 65.8% to 69.7% | 1,489 of 2,198 |
| Reading | 89.6% | 87.1% to 91.7% | 588 of 656 |
| Grammar | 58.4% | 56.0% to 60.9% | 901 of 1,542 |
Comparison group: pqr. Compare with another setup.
Accuracy by topic
Page images (preliminary)
Preliminary: these numbers are not final and may change. This setup is condition A of the Page images tab, which asks the 763 questions with underlined or numbered words as text (A), as a page image (B), as both (C) and through OCR models (D); the share answered right, with its 95% interval.
| Model | A · text onlyaccuracy, higher is better | B · image onlyaccuracy, higher is better | C · text plus imageaccuracy, higher is better | Best OCR routeaccuracy, higher is better |
|---|---|---|---|---|
| Drex Max v11 question per request | 64.1%60.6% to 67.4% | 46.8%43.3% to 50.3% | 66.1%62.6% to 69.3% | Not available: not run |
Confidence
- Answers kept at 99% or more
- 443, of which 1 wrong (share right: 98.7% to 100.0%, 95% interval); 1,755 skipped, the model less than 99% sure
- Calibration gap
- 6.3 points
- AUROC, all questions
- 0.83 0.81 to 0.84
- AUROC, reading
- 0.91 0.88 to 0.93
- AUROC, grammar
- 0.76 0.74 to 0.78
- Brier score
- 0.163
- Accuracy of its surest answers
- most confident fifth: 99.8% right; most confident half: 89.6% right; most confident four fifths: 76.5% right
- Typical score of a right answer
- 94.1%
- Typical score of a wrong answer
- 52.2%
Repeats
Run 5 times on all questions.
- Answers changed in 5 runs
- 0 of 2,198
- Same wrong answer in all 5
- 709 of 2,198
- Wrong in at least one run
- 709 of 2,198
- Accuracy in each run
- 67.7%, 67.7%, 67.7%, 67.7%, 67.7%
Latency
- Latency, median request
- 0.082 s (this run)
- Latency, 90th percentile
- 0.110 s
- Requests at once
- 1
- Throughput
- 673.5 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
- $0.023
- Whole run
- $0.051
- Basis
- Estimated at Nace.AI's list price for the same model; our hardware cost is not measured.
- Price
- $0.06 input, output free per 1M tokens (source, 2026-10-30)
- Input tokens per question
- 388
- 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 | 16.7% | −0.09 points | 16%§ |
| GPT-6 Astrahigh reasoning · 16 questions per request | 19.0% | −0.08 points | 19%§ |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 20.6% | −0.11 points | 19%§ |
| GPT-6 Astralow reasoning · 16 questions per request | 19.1% | −0.08 points | 19%§ |
| GPT-6 Astralow reasoning · 1 question per request | 16.7% | −0.09 points | 17%§ |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 16.9% | −0.10 points | 15%§ |
| GPT-6.1 Sollow reasoning · 1 question per request | 21.0% | −0.11 points | 21%§ |
| GPT-6.1 Sollow reasoning · 16 questions per request | 19.6% | −0.09 points | 18%§ |
| Claude Opus 5.5high reasoning · 16 questions per request | 16.7% | −0.09 points | 16%§ |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 19.1% | −0.09 points | 15%§ |
| Claude Opus 5.5low reasoning · 1 question per request | 20.9% | −0.10 points | 21%§ |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 16.7% | −0.09 points | 8%§ |
| GPT-6.1 Terralow reasoning · 16 questions per request | 20.7% | −0.14 points | 15%§ |
| Claude Opus 5.5low reasoning · 16 questions per request | 20.1% | −0.14 points | 20%§ |
| Claude Sonnet 5.5high reasoning · 1 question per request | 24.1% | −0.17 points | 24%§ |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 48.0% | −0.04 points | 47%§ |
§ Counts the first model at its maker’s list price for the same model; our hardware cost is not measured.