Calibration
Terminal-Bench 2.0 → Terminal-Bench Hard
Terminal-Bench Hard is estimated from Terminal-Bench 2.0 with a Hill curve fitted on 12 models measured on both: y = 0.4107 + (1.2000 − 0.4107)·x^6.00 / (0.99545^6.00 + x^6.00), R² = 0.62, cross-validated error 4.0 pp. It is used for 18 estimates.
| Estimated model | Terminal-Bench 2.0 | Terminal-Bench Hard | Source |
|---|---|---|---|
| GPT-5.3 Codex | 77.3% | 55.3% | estimated ± 4.0 pp, high confidence |
| Claude Opus 4.7 (Adaptive) | 69.4% | 49.2% | estimated ± 4.0 pp, high confidence |
| Composer 2 | 61.7% | 45.3% | estimated ± 4.0 pp, high confidence |
| Claude Opus 4.5 | 59.3% | 44.4% | estimated ± 4.0 pp, high confidence |
| Qwen3.6-27B | 59.3% | 44.4% | estimated ± 4.0 pp, high confidence |
| Hy3 Preview | 54.4% | 43.1% | estimated ± 4.0 pp, medium confidence |
| Qwen3.5 397B | 52.5% | 42.7% | estimated ± 4.0 pp, medium confidence |
| Qwen3.6-35B-A3B | 51.5% | 42.6% | estimated ± 4.0 pp, medium confidence |
| Kimi K2.5 | 50.8% | 42.4% | estimated ± 4.0 pp, medium confidence |
| Kimi K2.5 (Reasoning) | 50.8% | 42.4% | estimated ± 4.0 pp, medium confidence |
| Claude Sonnet 4.5 | 50.0% | 42.3% | estimated ± 4.0 pp, medium confidence |
| Qwen3.5-122B-A10B | 49.4% | 42.2% | estimated ± 4.0 pp, medium confidence |
| MAI-Thinking-1 | 46.0% | 41.8% | estimated ± 4.0 pp, medium confidence |
| Qwen3.5-27B | 41.6% | 41.5% | estimated ± 4.0 pp, medium confidence |
| GLM-4.7 | 41.0% | 41.5% | estimated ± 4.0 pp, medium confidence |
| Qwen3.5-35B-A3B | 40.5% | 41.4% | estimated ± 4.0 pp, medium confidence |
| Laguna XS 2.1 | 37.5% | 41.3% | estimated ± 4.0 pp, medium confidence |
| LongCat-Flash-Lite-Sparse | 33.7% | 41.2% | estimated ± 4.0 pp, medium confidence |
Fitted on
| Model | Terminal-Bench 2.0 | Terminal-Bench Hard |
|---|---|---|
| Claude Opus 4.6 | 65.4% | 48.5% |
| GLM-5 | 56.2% | 43.2% |
| GPT-5.4 | 75.1% | 57.6% |
| GPT-5.4 mini | 60.0% | 52.3% |
| GPT-5.5 | 82.0% | 60.6% |
| Kimi K2.6 | 66.7% | 43.9% |
| MiMo-V2.5-Pro | 68.4% | 43.2% |
| Muse Spark | 59.0% | 45.5% |
| Qwen 3.6 Max (preview) | 65.4% | 43.9% |
| Qwen3.6 Plus | 61.6% | 43.9% |
| Qwen3.7 Max | 69.7% | 50.8% |
| Qwen3.7 Plus | 70.3% | 47.0% |