Calibration
Vals LiveCodeBench → CursorBench 3.2
CursorBench 3.2 is estimated from Vals LiveCodeBench with a offset logistic curve fitted on 16 models measured on both: y = 0.5910 + (0.7302 − 0.5910) / (1 + exp(−200.00·(x − 0.8875))), R² = 0.40, cross-validated error 7.2 pp. It is used for 5 estimates.
| Estimated model | Vals LiveCodeBench | CursorBench 3.2 | Source |
|---|---|---|---|
| GPT-5.1-Codex | 85.6% | 59.1% | estimated ± 7.2 pp, low confidence |
| GPT-5.1-Codex-Max | 83.6% | 59.1% | estimated ± 7.2 pp, low confidence |
| GLM-4.6 | 81.0% | 59.1% | estimated ± 7.2 pp, low confidence |
| GLM-4.5 | 67.4% | 59.1% | estimated ± 7.2 pp, low confidence |
| Grok Code Fast 1 | 62.0% | 59.1% | estimated ± 7.2 pp, low confidence |
Fitted on
| Model | Vals LiveCodeBench | CursorBench 3.2 |
|---|---|---|
| Claude Fable 5 | 89.8% | 70.5% |
| Claude Fable 5.1 | 90.5% | 73.4% |
| Claude Opus 4.8 | 87.8% | 62.3% |
| Claude Opus 5 | 89.0% | 70.0% |
| Claude Sonnet 5 | 82.4% | 61.5% |
| Gemini 3.5 Flash | 87.6% | 48.8% |
| Gemini 3.6 Flash | 88.1% | 53.5% |
| Gemini 3.8 Flash | 89.5% | 69.2% |
| GLM-5.2 | 69.5% | 55.0% |
| GPT-5.5 | 85.3% | 58.4% |
| GPT-5.6 Sol | 82.6% | 67.2% |
| GPT-5.6 Terra | 85.9% | 64.9% |
| Grok 4.5 | 87.4% | 66.7% |
| Grok 4.6 | 88.2% | 70.8% |
| Kimi K2.7 Code | 82.1% | 49.7% |
| Kimi K3 | 87.2% | 60.8% |