Vals AI Terminal-Bench Science: leaderboard
Measures long-horizon scientific work in a terminal across 70 tasks spanning life, physical, mathematical, earth and engineering sciences. Accuracy is the percentage of tasks that pass their verifier suite, using the overall result reported by Vals AI.
Metric: Resolution Rate (%). Source: www.vals.ai. Saturation forecast: Around June 2027. 43 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Max) | 62.86 | #6 (GPT-6 Astra) |
| 2 | GPT-6.1 Sol (Max) | 52.86 | #7 (GPT-6.1 Sol) |
| 3 | Claude Opus 5.5 (Max) | 47.14 | #3 (Claude Opus 5.5) |
| 4 | Claude Sonnet 5.5 (Max) | 47.14 | #14 (Claude Sonnet 5.5) |
| 5 | Gemini 4 Argon (High) | 44.29 | #9 (Gemini 4 Argon) |
| 6 | Claude Fable 5.1 (Max) | 40 | #10 (Claude Fable 5.1) |
| 7 | Muse Wonder Full (Max) | 37.14 | #12 (Muse Wonder Full) |
| 8 | Kimi K3.1 Pro (Max) | 31.43 | #17 (Kimi K3.1 Pro) |
| 9 | GPT-6 Sol (Max) | 30 | #19 (GPT-6 Sol) |
| 10 | Claude Opus 5 (Max) | 27.14 | #16 (Claude Opus 5) |
| 11 | Muse Wonder (Max) | 22.86 | #18 (Muse Wonder) |
| 12 | GPT-5.6 Sol (Max) | 20 | #22 (GPT-5.6 Sol) |
| 13 | Claude Fable 5 (Max) | 15.71 | #20 (Claude Fable 5) |
| 14 | Qwen 3.8 Max | 12.86 | #34 |
| 15 | Grok 4.6 (High) | 11.43 | #29 (Grok 4.6) |
Interactive version: theaggregate.ai/benchmark?slug=vals-ai-terminal-bench-science · How It Works · Data refreshed daily, snapshot 2026-11-05.