Calibration
Terminal-Bench 2.1 (Vals AI) → Terminal-Bench Hard
Terminal-Bench Hard is estimated from Terminal-Bench 2.1 (Vals AI) with a Hill curve fitted on 14 models measured on both: y = 0.4319 + (0.9323 − 0.4319)·x^6.00 / (0.87197^6.00 + x^6.00), R² = 0.81, cross-validated error 3.8 pp. It is used for 46 estimates.
| Estimated model | Terminal-Bench 2.1 (Vals AI) | Terminal-Bench Hard | Source |
|---|---|---|---|
| GPT-6 Astra | 87.3% | 68.3% | estimated ± 3.8 pp, medium confidence |
| Claude Fable 5.1 | 85.0% | 66.3% | estimated ± 3.8 pp, high confidence |
| Claude Opus 5 | 84.6% | 65.9% | estimated ± 3.8 pp, high confidence |
| Gemini 3.8 Flash | 81.3% | 63.0% | estimated ± 3.8 pp, high confidence |
| Kimi K3 | 80.9% | 62.7% | estimated ± 3.8 pp, high confidence |
| GPT-5.6 Luna | 79.0% | 61.0% | estimated ± 3.8 pp, high confidence |
| Grok 4.6 | 78.3% | 60.4% | estimated ± 3.8 pp, high confidence |
| Gemini 3.7 Flash | 77.5% | 59.7% | estimated ± 3.8 pp, high confidence |
| GPT-5.6 Terra | 77.5% | 59.7% | estimated ± 3.8 pp, high confidence |
| Claude Sonnet 5 | 74.5% | 57.2% | estimated ± 3.8 pp, high confidence |
| DeepSeek V4.1 Flash 1017 | 74.5% | 57.2% | estimated ± 3.8 pp, high confidence |
| Gemini 3.5 Flash | 74.2% | 57.0% | estimated ± 3.8 pp, high confidence |
| Gemini 3.6 Flash | 73.8% | 56.6% | estimated ± 3.8 pp, high confidence |
| Grok 4.7 | 73.4% | 56.3% | estimated ± 3.8 pp, high confidence |
| Muse Spark 1.3 | 72.3% | 55.5% | estimated ± 3.8 pp, high confidence |
| GLM-5.3 | 71.5% | 54.9% | estimated ± 3.8 pp, high confidence |
| Muse Spark 1.2 | 69.7% | 53.5% | estimated ± 3.8 pp, high confidence |
| Muse Spark 1.1 | 69.3% | 53.3% | estimated ± 3.8 pp, high confidence |
| Grok 4.5 | 67.8% | 52.2% | estimated ± 3.8 pp, high confidence |
| Qwen3.8 Max | 67.4% | 52.0% | estimated ± 3.8 pp, high confidence |
| DeepSeek V4 Flash 0731 | 67.0% | 51.7% | estimated ± 3.8 pp, high confidence |
| GLM-5.3-Flash | 62.9% | 49.4% | estimated ± 3.8 pp, high confidence |
| MiMo-V2.5 | 60.7% | 48.3% | estimated ± 3.8 pp, high confidence |
| Composer 2.5 | 58.4% | 47.3% | estimated ± 3.8 pp, high confidence |
| Qwen3.8-27B | 58.4% | 47.3% | estimated ± 3.8 pp, high confidence |
| Claude Sonnet 4.6 | 57.3% | 46.9% | estimated ± 3.8 pp, high confidence |
| GLM-5.1 | 56.9% | 46.8% | estimated ± 3.8 pp, high confidence |
| Hy4 preview | 55.1% | 46.2% | estimated ± 3.8 pp, high confidence |
| Inkling-Small | 55.1% | 46.2% | estimated ± 3.8 pp, high confidence |
| DeepSeek V4 Pro 0813 | 54.7% | 46.1% | estimated ± 3.8 pp, high confidence |
| Gemini 3 Flash | 53.9% | 45.8% | estimated ± 3.8 pp, high confidence |
| MiniMax M3 | 53.6% | 45.7% | estimated ± 3.8 pp, high confidence |
| Nemotron 3 Ultra | 50.9% | 45.1% | estimated ± 3.8 pp, medium confidence |
| Gemini 3.5 Flash-Lite | 50.2% | 44.9% | estimated ± 3.8 pp, medium confidence |
| Ling 3.0 Flash | 50.2% | 44.9% | estimated ± 3.8 pp, medium confidence |
| MiniMax M2.7 | 48.7% | 44.7% | estimated ± 3.8 pp, medium confidence |
| Inkling | 47.6% | 44.5% | estimated ± 3.8 pp, medium confidence |
| Grok 4.20 | 44.2% | 44.0% | estimated ± 3.8 pp, medium confidence |
| Claude Haiku 4.5 | 43.8% | 44.0% | estimated ± 3.8 pp, medium confidence |
| Grok 4.3 | 41.9% | 43.8% | estimated ± 3.8 pp, medium confidence |
| GPT-5.4 nano | 41.6% | 43.8% | estimated ± 3.8 pp, medium confidence |
| Mistral Medium 3.5 128B | 39.0% | 43.6% | estimated ± 3.8 pp, medium confidence |
| Mercury 2.5 | 34.5% | 43.4% | estimated ± 3.8 pp, medium confidence |
| Gemini 3.1 Flash-Lite | 34.1% | 43.4% | estimated ± 3.8 pp, medium confidence |
| Laguna M.1 | 34.1% | 43.4% | estimated ± 3.8 pp, medium confidence |
| Laguna XS.2 | 25.8% | 43.2% | estimated ± 3.8 pp, medium confidence |
Fitted on
| Model | Terminal-Bench 2.1 (Vals AI) | Terminal-Bench Hard |
|---|---|---|
| Claude Fable 5 | 80.5% | 62.9% |
| Claude Opus 4.7 | 68.5% | 54.5% |
| Claude Opus 4.8 | 71.9% | 58.3% |
| Gemini 3.1 Pro | 70.8% | 53.8% |
| GLM-5.2 | 67.8% | 50.8% |
| GPT-5.4 mini | 54.7% | 52.3% |
| GPT-5.5 | 76.4% | 60.6% |
| GPT-5.6 Sol | 85.8% | 65.9% |
| Kimi K2.6 | 53.6% | 43.9% |
| Kimi K2.7 Code | 67.0% | 44.7% |
| MiMo-V2.5-Pro | 57.3% | 43.2% |
| Qwen3.6 Plus | 53.2% | 43.9% |
| Qwen3.7 Max | 61.0% | 50.8% |
| Qwen3.7 Plus | 52.8% | 47.0% |