Calibration
Terminal-Bench 4.1 → Terminal-Bench 2.1 (Vals AI)
Terminal-Bench 2.1 (Vals AI) is estimated from Terminal-Bench 4.1 with a Michaelis–Menten + offset curve fitted on 32 models measured on both: y = 0.5020 + 0.3353·x / (0.07787 + x), R² = 0.52, cross-validated error 8.3 pp. It is used for 2 estimates.
| Estimated model | Terminal-Bench 4.1 | Terminal-Bench 2.1 (Vals AI) | Source |
|---|---|---|---|
| GPT-6.1 Sol | 60.9% | 79.9% | estimated ± 8.3 pp, medium confidence |
| Hy4 | 37.9% | 78.0% | estimated ± 8.3 pp, medium confidence |
Fitted on
| Model | Terminal-Bench 4.1 | Terminal-Bench 2.1 (Vals AI) |
|---|---|---|
| Claude Fable 5 | 44.9% | 80.5% |
| Claude Fable 5.1 | 57.6% | 85.0% |
| Claude Opus 4.8 | 23.6% | 71.9% |
| Claude Opus 5 | 53.3% | 84.6% |
| Claude Opus 5.5 | 65.8% | 87.6% |
| Claude Sonnet 5 | 12.1% | 74.5% |
| Claude Sonnet 5.5 | 63.9% | 83.1% |
| DeepSeek V4.1 Flash 1017 | 38.5% | 74.5% |
| DeepSeek V4 Pro | 6.7% | 54.7% |
| Gemini 3.1 Pro | 2.7% | 70.8% |
| Gemini 3.7 Flash | 12.4% | 77.5% |
| Gemini 3.8 Flash | 18.2% | 81.3% |
| Gemini 4 Argon | 55.5% | 57.7% |
| GLM-5.3 | 41.2% | 71.5% |
| GLM-5.3-Flash | 37.0% | 62.9% |
| GPT-5.4 mini | 1.8% | 54.7% |
| GPT-5.6 Luna | 17.6% | 79.0% |
| GPT-5.6 Sol | 38.5% | 85.8% |
| GPT-5.6 Terra | 21.5% | 77.5% |
| GPT-6 Astra | 60.3% | 87.3% |
| GPT-6 Luna | 15.2% | 73.0% |
| GPT-6 Sol | 50.3% | 83.1% |
| Grok 4.5 | 12.4% | 67.8% |
| Grok 4.6 | 19.7% | 78.3% |
| Grok 4.7 | 38.8% | 73.4% |
| Inkling | 0.0% | 47.6% |
| Inkling-Small | 0.9% | 55.1% |
| Kimi K3 | 23.0% | 80.9% |
| MiMo-V2.6-Flash | 16.1% | 76.4% |
| MiMo-V2.6-Pro | 27.0% | 67.8% |
| MiniMax M3 | 7.6% | 53.6% |
| Muse Spark 1.3 | 14.9% | 72.3% |