Calibration
Terminal-Bench Hard → Terminal-Bench 2.0
Terminal-Bench 2.0 is estimated from Terminal-Bench Hard with a Hill curve fitted on 12 models measured on both: y = 0.5905 + (1.2000 − 0.5905)·x^6.00 / (0.68474^6.00 + x^6.00), R² = 0.57, cross-validated error 5.4 pp. It is used for 11 estimates.
| Estimated model | Terminal-Bench Hard | Terminal-Bench 2.0 | Source |
|---|---|---|---|
| GPT-5.6 Sol | 65.9% | 86.0% | estimated ± 5.4 pp, low confidence |
| Claude Fable 5 | 62.9% | 81.9% | estimated ± 5.4 pp, low confidence |
| Claude Opus 4.8 | 58.3% | 75.9% | estimated ± 5.4 pp, medium confidence |
| Claude Opus 4.7 | 54.5% | 71.4% | estimated ± 5.4 pp, medium confidence |
| Gemini 3.1 Pro | 53.8% | 70.7% | estimated ± 5.4 pp, medium confidence |
| GLM-5.2 | 50.8% | 67.8% | estimated ± 5.4 pp, medium confidence |
| Claude Opus 4.5 Thinking | 47.0% | 64.8% | estimated ± 5.4 pp, medium confidence |
| GPT-5.2 | 47.0% | 64.8% | estimated ± 5.4 pp, medium confidence |
| GPT-5.1 | 45.5% | 63.9% | estimated ± 5.4 pp, medium confidence |
| Kimi K2.7 Code | 44.7% | 63.4% | estimated ± 5.4 pp, medium confidence |
| Solar Open 2 | 28.3% | 59.3% | estimated ± 5.4 pp, low confidence |
Fitted on
| Model | Terminal-Bench Hard | Terminal-Bench 2.0 |
|---|---|---|
| Claude Opus 4.6 | 48.5% | 65.4% |
| GLM-5 | 43.2% | 56.2% |
| GPT-5.4 | 57.6% | 75.1% |
| GPT-5.4 mini | 52.3% | 60.0% |
| GPT-5.5 | 60.6% | 82.0% |
| Kimi K2.6 | 43.9% | 66.7% |
| MiMo-V2.5-Pro | 43.2% | 68.4% |
| Muse Spark | 45.5% | 59.0% |
| Qwen 3.6 Max (preview) | 43.9% | 65.4% |
| Qwen3.6 Plus | 43.9% | 61.6% |
| Qwen3.7 Max | 50.8% | 69.7% |
| Qwen3.7 Plus | 47.0% | 70.3% |