benchgap
Calibration

FrontierCode 2.0 → CursorBench 3.2

CursorBench 3.2 is estimated from FrontierCode 2.0 with a Hill curve fitted on 6 models measured on both: y = 0.5996 + (0.8690 − 0.5996)·x^6.00 / (0.33655^6.00 + x^6.00), R² = 0.79, cross-validated error 4.1 pp. It is used for 6 estimates.

Estimated modelFrontierCode 2.0CursorBench 3.2Source
GPT-6.2 Astra44.2%82.5%estimated ± 4.1 pp, low confidence
GPT-6.2 Sol43.8%82.3%estimated ± 4.1 pp, low confidence
Inkling 2 Small22.6%62.2%estimated ± 4.1 pp, medium confidence
Kimi K3.1 Pro19.9%61.1%estimated ± 4.1 pp, medium confidence
Mistral Large 419.7%61.0%estimated ± 4.1 pp, medium confidence
Muse Spark Open6.5%60.0%estimated ± 4.1 pp, low confidence

Fitted on

ModelFrontierCode 2.0CursorBench 3.2
Claude Fable 5.134.0%73.4%
Claude Opus 4.815.3%62.3%
Claude Opus 528.1%70.0%
GPT-5.523.6%58.4%
GPT-5.6 Sol28.9%67.2%
Kimi K318.9%60.8%