Calibration
VulcanBench v3 → CursorBench 4.0
CursorBench 4.0 is estimated from VulcanBench v3 with a offset logistic curve fitted on 5 models measured on both: y = 0.3507 + (0.4866 − 0.3507) / (1 + exp(−199.55·(x − 0.8687))), R² = 0.65, cross-validated error 3.6 pp. It is used for 1 estimate.
| Estimated model | VulcanBench v3 | CursorBench 4.0 | Source |
|---|---|---|---|
| Claude Haiku 4.5 | 76.2% | 35.1% | estimated ± 3.6 pp, low confidence |