Benchmarks / Agentic
Terminal-Bench Science (Vals)
Research workflow tasks contributed by practicing scientists across five domains
Terminal-Bench Science (Vals)
Higher is better · best effort per model
full table below
| # | Model | Effort | Score |
|---|---|---|---|
| 1 | GPT-6 Astra | max | 62.9% |
| 2 | GPT-6.1 Sol | max | 52.9% |
| 3 | Opus 5.5 | max | 47.1% |
| 4 | Sonnet 5.5 | max | 47.1% |
| 5 | Gemini 4 | high | 44.3% |
| 6 | Fable 5.1 | max | 40.0% |
| 7 | Wonder Full | max | 37.1% |
| 8 | GPT-6 Sol | max | 30.0% |
| 9 | Opus 5 | max | 27.1% |
| 10 | Wonder | max | 22.9% |
| 11 | GPT-5.6 Sol | max | 20.0% |
| 12 | Fable 5 | max | 15.7% |
| 13 | Qwen 3.8 Max | max | 12.9% |
| 14 | Grok 4.6 | high | 11.4% |
| 15 | Muse Spark 1.3 | max | 10.0% |
| 16 | Grok 4.7 | xhigh | 10.0% |
| 17 | Haiku 5.5 | max | 10.0% |
| 18 | GPT-5.6 Terra | max | 10.0% |
| 19 | Gemini 3.8 Flash | high | 8.6% |
| 20 | Gemini 3.7 Flash | high | 7.1% |
| 21 | Gemini 3.5 Flash | high | 5.7% |
| 22 | GLM 5.3 | max | 5.7% |
| 23 | Sonnet 5 | max | 5.7% |
| 24 | Ember-1 | – | 5.7% |
| 25 | GPT-6 Luna | max | 4.3% |
| 26 | DeepSeek V4 Flash 0731 | high | 4.3% |
| 27 | Gemini 3.6 Flash | high | 4.3% |
| 28 | Opus 4.8 | max | 4.3% |
| 29 | DeepSeek V4 Pro 0813 | max | 4.3% |
| 30 | MiMo V2.6 Flash | high | 4.3% |
| 31 | Step 5 Preview | – | 2.9% |
| 32 | MiMo V2.6 Pro | high | 2.9% |
| 33 | GLM 5.3 Flash | max | 2.9% |
| 34 | Gemini 3.1 Pro | high | 1.4% |
| 35 | Hy4 | – | 1.4% |
| 36 | Kimi K3 | max | 1.4% |
| 37 | Mercury 2.5 | high | 0.0% |
| 38 | GPT-5.6 Luna | max | 0.0% |
| 39 | DeepSeek V4.1 Flash | high | 0.0% |
| 40 | Qwen 3.8 27B | xhigh | 0.0% |