Terminal-Bench Science 0.1 (Vals)
Science · 2026-09-23
Seventy expert-authored scientific terminal tasks evaluated with Terminus 2 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 |
| Claude Opus 5.5 | 48.6 |
| Claude Fable 5.1 | 34.3 |
| Claude Opus 5 | 22.9 |
| Muse Spark 1.3 | 14.3 |
| Claude Fable 5 | 12.9 |
| GPT-5.6 Sol | 12.9 |
| Grok 4.7 | 11.4 |
| Opus 4.8 | 10.0 |
| GPT-5.6 Terra | 10.0 |
| Gemini 3.8 Flash | 8.6 |
| Gemini 3.7 Flash | 5.7 |
| GPT-5.6 Luna | 5.7 |
| MiMo-V2.6-Flash | 5.7 |
| DeepSeek V4.1 Flash | 4.3 |
| Gemini 3.5 Flash | 4.3 |
| GLM-5.3 | 4.3 |
| Grok 4.6 | 4.3 |
| Sonnet 5 | 2.9 |
| Kimi K3 | 2.9 |
| MiMo-V2.6-Pro | 2.9 |
| Gemini 3.6 Flash | 1.4 |
| Hy4 Preview | 1.4 |
| Qwen3.8 27B | 1.4 |
| DeepSeek-V4-Flash-0731 | 0.0 |
| DeepSeek V4 Pro 0813 | 0.0 |
| GLM-5.3 Flash | 0.0 |
| Mercury 2.5 | 0.0 |