Calibration

CursorBench 3.2 → SWE-bench Pro

Name: benchgap: LLM benchmark scores, measured and estimated
Creator: benchgap

SWE-bench Pro is estimated from CursorBench 3.2 with a Hill curve fitted on 12 models measured on both: y = 0.5038 + (1.2000 − 0.5038)·x^6.00 / (0.77813^6.00 + x^6.00), R² = 0.78, cross-validated error 4.5 pp. It is used for 3 estimates.

Estimated model	CursorBench 3.2	SWE-bench Pro	Source
Gemini 3.6 Flash	53.5%	57.0%	estimated ± 4.5 pp, high confidence
Kimi K2.7 Code	49.7%	54.8%	estimated ± 4.5 pp, high confidence
Kimi K3	60.8%	63.3%	estimated ± 4.5 pp, high confidence