Decision model
Strands Decider 2B
1 question per request
27.5% of the 2,198 questions right (25.7% to 29.4%, 95% interval). Back to the leaderboard.
How we called it
- Kind
- Decision model
- Access
- Strands Decider server, self-hosted
- Paid as
- Our own hardware
- Model ID
strands-decider-2b- Parameters
- 2.3B
- Hardware
- 1 × RTX PRO 6000 Blackwell (96 GB); BF16
- Note
- Amazon's decision model: a LoRA adapter and a decision head on Qwen3.5-2B-Base
- Confidence from
- Option score
- Run
- 141, 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 | 27.5% | 25.7% to 29.4% | 604 of 2,198 |
| Reading | 36.9% | 33.3% to 40.6% | 242 of 656 |
| Grammar | 23.5% | 21.4% to 25.7% | 362 of 1,542 |
Comparison group: G. 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
- 12.2 points
- AUROC, all questions
- 0.65 0.62 to 0.67
- AUROC, reading
- 0.74 0.70 to 0.78
- AUROC, grammar
- 0.57 0.54 to 0.61
- Brier score
- 0.198
- Accuracy of its surest answers
- most confident fifth: 48.7% right; most confident half: 35.9% right; most confident four fifths: 29.4% right
- Typical score of a right answer
- 40.6%
- Typical score of a wrong answer
- 33.5%
Repeats
We did not repeat this setup.
Latency
- Latency, median request
- 0.026 s (this run)
- Latency, 90th percentile
- 0.030 s
- Requests at once
- 1
- Throughput
- 2062.5 questions per minute of request time
- Request time
- 1.1 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
- 359
- Output tokens per question
- 1
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 | 0.8% | −0.09 points | 1%† |
| GPT-6 Astrahigh reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| GPT-6.1 Solhigh reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| GPT-6 Astralow reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| GPT-6 Astralow reasoning · 1 question per request | 0.8% | −0.09 points | 1%† |
| Gemini 3.8 Flashlow reasoning · 1 question per request | 0.8% | −0.09 points | 1%† |
| GPT-6.1 Sollow reasoning · 1 question per request | 0.8% | −0.09 points | 1%† |
| GPT-6.1 Sollow reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| Claude Opus 5.5high reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| GPT-6.1 Terrahigh reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| Claude Opus 5.5low reasoning · 1 question per request | 0.8% | −0.09 points | 1%† |
| Gemini 3.8 Flashlow reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| GPT-6.1 Terralow reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| Claude Opus 5.5low reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
| Claude Sonnet 5.5high reasoning · 1 question per request | 0.8% | −0.09 points | 1%† |
| Claude Sonnet 5.5high reasoning · 16 questions per request | 0.8% | −0.09 points | 1%† |
† Leaves out the first model’s hardware cost, which we have not estimated.