benchgap
Calibration

CursorBench 3.2 → ProgramBench

ProgramBench is estimated from CursorBench 3.2 with a Hill curve fitted on 5 models measured on both: y = 0.1578 + (1.0080 − 0.1578)·x^6.00 / (0.51881^6.00 + x^6.00), R² = 0.97, cross-validated error 6.0 pp. It is used for 9 estimates.

Estimated modelCursorBench 3.2ProgramBenchSource
Claude Fable 570.5%89.1%estimated ± 6.0 pp, low confidence
Claude Sonnet 561.5%78.3%estimated ± 6.0 pp, low confidence
Composer 2.556.1%68.1%estimated ± 6.0 pp, low confidence
Gemini 3.5 Flash48.8%50.6%estimated ± 6.0 pp, low confidence
Gemini 3.6 Flash53.5%62.2%estimated ± 6.0 pp, low confidence
Gemini 3.8 Flash69.2%88.0%estimated ± 6.0 pp, low confidence
GPT-5.6 Luna61.1%77.6%estimated ± 6.0 pp, low confidence
GPT-5.6 Terra64.9%83.2%estimated ± 6.0 pp, low confidence
Grok 4.670.8%89.4%estimated ± 6.0 pp, low confidence