Calibration
Terminal-Bench 3.0 → Terminal-Bench 2.1 (Vals AI)
Terminal-Bench 2.1 (Vals AI) is estimated from Terminal-Bench 3.0 with a linear curve fitted on 11 models measured on both: y = 0.4268·x + 0.6738, R² = 0.65, cross-validated error 4.0 pp. It is used for 0 estimates.
Fitted on
| Model | Terminal-Bench 3.0 | Terminal-Bench 2.1 (Vals AI) |
|---|---|---|
| Claude Fable 5 | 34.1% | 80.5% |
| Claude Opus 4.8 | 21.1% | 71.9% |
| Claude Opus 5 | 42.7% | 84.6% |
| Claude Sonnet 5 | 14.6% | 74.5% |
| Gemini 3.7 Flash | 14.9% | 77.5% |
| GLM-5.2 | 4.6% | 67.8% |
| GPT-5.6 Luna | 14.3% | 79.0% |
| GPT-5.6 Sol | 34.6% | 85.8% |
| GPT-5.6 Terra | 20.8% | 77.5% |
| Grok 4.5 | 15.7% | 67.8% |
| Grok 4.6 | 26.5% | 78.3% |