AstaBench
Can the model reason accurately about difficult scientific material?
Science7 ranked models7 reported values1 reportsHigher is better
Top rankings
One row per model, using its best reported score across effort settings.
| Rank | Model | Best score | Best reported setting |
|---|---|---|---|
| 1 | Claude Opus 4.7Anthropic | 58.0% | max effort$3.54 / task1 value · 1 reportApr 30, 2026 · source |
| 2 | Claude Opus 4.6Anthropic | 55.3% | max effortAllen AI report1 value · 1 reportApr 30, 2026 · source |
| 3 | Claude Sonnet 4.6Anthropic | 54.5% | max effortAllen AI report1 value · 1 reportApr 30, 2026 · source |
| 4 | Asta v0Allen AI | 53.0% | Reported configurationAllen AI report1 value · 1 reportApr 30, 2026 · source |
| 5 | GPT-5.5OpenAI | 52.9% | xhigh effort$1.61 / task1 value · 1 reportApr 30, 2026 · source |
| 6 | Gemini 3.1 Pro PreviewGoogle | 49.6% | high effortAllen AI report1 value · 1 reportApr 30, 2026 · source |
| 7 | GPT-5.4OpenAI | 46.5% | xhigh effortAllen AI report1 value · 1 reportApr 30, 2026 · source |
Effort curve
Every sourced cost-linked effort value for this exact version. Lines connect complete sweeps only.
No cost-linked effort sweep for this version.
Definition and comparison boundarypublic methodology
Can the model reason accurately about difficult scientific material?
Scientific knowledge and reasoning under the benchmark's published question set. Higher is better. The value is the percentage reported in this lab's table.
Harness, tools, prompts, attempt count, and benchmark version can materially change the result. Compare within one reporting context.
ReAct agent framework. The source directly reports observed per-problem cost only for Claude Opus 4.7 ($3.54) and GPT-5.5 ($1.61).
Method / source