benchgap
Calibration

CWE-bench v1 → CWE-bench

CWE-bench is estimated from CWE-bench v1 with a Michaelis–Menten curve fitted on 6 models measured on both: y = 2.0000·x / (1.49229 + x), R² = 0.52, cross-validated error 6.6 pp. It is used for 2 estimates.

Estimated modelCWE-bench v1CWE-benchSource
Grok 4.657.0%55.3%estimated ± 6.6 pp, low confidence
Muse Spark 1.355.0%53.9%estimated ± 6.6 pp, low confidence

Fitted on

ModelCWE-bench v1CWE-bench
Claude Opus 5.567.0%67.1%
DeepSeek V4.1 Flash 101755.0%57.6%
Gemini 4 Argon68.0%65.0%
GPT-6 Astra68.0%64.2%
Grok 4.768.0%54.7%
Hy4 preview53.0%47.3%