Terminal-Bench Science 0.1 (Vals)
Science · 2026-09-23
Seventy expert-authored scientific terminal tasks evaluated with the row-disclosed Terminus 2 or Mini-SWE-Agent harness and an eight-hour agent budget. One graded result per task; infrastructure/provider errors are retried, including some successful reattempts, while graded failures are not. This retry-conditioned protocol is separate from the official native-agent three-trial evaluation.
Top models (higher is better)
| Model | Score |
|---|---|
| GPT-6 Astra | 65.7 |
| GPT-6.1 Sol | 52.9 |
| Claude Opus 5.5 | 48.6 |
| Claude Sonnet 5.5 | 47.1 |
| Gemini 4 Argon | 44.3 |
| Muse Wonder Full | 37.1 |
| Claude Fable 5.1 | 34.3 |
| GPT-6 Sol | 31.4 |
| Kimi K3.1 Pro | 31.4 |
| Claude Opus 5 | 27.1 |
| Muse Wonder | 22.9 |
| GPT-5.6 Sol | 20.0 |
| Muse Spark 1.3 | 14.3 |
| Claude Fable 5 | 12.9 |
| Grok 4.6 | 11.4 |
| Grok 4.7 | 11.4 |
| Claude Haiku 5.5 | 10.0 |
| Opus 4.8 | 10.0 |
| GPT-5.6 Terra | 10.0 |
| GPT-6 Luna | 10.0 |
| Gemini 3.8 Flash | 8.6 |
| Gemini 3.7 Flash | 7.1 |
| Sonnet 5 | 5.7 |
| Ember-1 | 5.7 |
| Gemini 3.5 Flash | 5.7 |
| GLM-5.3 | 5.7 |
| GPT-5.6 Luna | 5.7 |
| MiMo-V2.6-Flash | 5.7 |
| DeepSeek V4.1 Flash | 4.3 |
| DeepSeek-V4-Flash-0731 | 4.3 |
| DeepSeek V4 Pro 0813 | 4.3 |
| Gemini 3.6 Flash | 4.3 |
| GLM-5.3 Flash | 2.9 |
| MiMo-V2.6-Pro | 2.9 |
| Step 5 Preview | 2.9 |
| Beam 1.0 501B A23B | 1.4 |
| Gemini 3.1 Pro Preview | 1.4 |
| Hy4 Preview | 1.4 |
| Kimi K3 | 1.4 |
| Mercury 2.5 | 0.0 |