benchgap
Calibration

CursorBench 4.0 → SWE-bench Pro

SWE-bench Pro is estimated from CursorBench 4.0 with a Hill curve fitted on 9 models measured on both: y = 0.5849 + (1.0973 − 0.5849)·x^6.00 / (0.54212^6.00 + x^6.00), R² = 0.93, cross-validated error 4.2 pp. It is used for 8 estimates.

Estimated modelCursorBench 4.0SWE-bench ProSource
Gemini 4 Carbon60.4%92.1%estimated ± 4.2 pp, medium confidence
Gemini 4 Hydrogen54.3%84.2%estimated ± 4.2 pp, high confidence
Muse Wonder50.4%78.6%estimated ± 4.2 pp, high confidence
Claude Haiku 5.548.4%75.7%estimated ± 4.2 pp, high confidence
Grok 4.746.3%72.8%estimated ± 4.2 pp, high confidence
Muse Spark 1.341.6%67.2%estimated ± 4.2 pp, high confidence
Grok 4.641.4%67.0%estimated ± 4.2 pp, high confidence
Composer 2.527.7%59.4%estimated ± 4.2 pp, medium confidence

Fitted on

ModelCursorBench 4.0SWE-bench Pro
Claude Fable 5.151.8%81.2%
Claude Fable 5.260.0%93.0%
Claude Opus 546.6%79.2%
Claude Opus 5.557.8%89.9%
Claude Sonnet 534.1%63.2%
Claude Sonnet 5.555.5%81.3%
GPT-5.6 Luna35.9%62.7%
GPT-5.6 Sol41.7%64.6%
GPT-5.6 Terra41.3%63.4%