benchgap
Calibration

CyberGym → ExploitGym

ExploitGym is estimated from CyberGym with a offset logistic curve fitted on 11 models measured on both: y = 0.0148 + (0.1783 − 0.0148) / (1 + exp(−77.87·(x − 0.7838))), R² = 0.36, cross-validated error 9.3 pp. It is used for 8 estimates.

Estimated modelCyberGymExploitGymSource
Claude Opus 4.550.6%1.5%estimated ± 9.3 pp, low confidence
Gemini 3.5 Flash Cyber83.2%17.5%estimated ± 9.3 pp, low confidence
Gemini 3.8 Flash Cyber86.2%17.8%estimated ± 9.3 pp, low confidence
GLM-543.2%1.5%estimated ± 9.3 pp, low confidence
Hy4 preview78.4%9.7%estimated ± 9.3 pp, low confidence
Ling 3.1 Flash87.9%17.8%estimated ± 9.3 pp, low confidence
Muse Spark43.5%1.5%estimated ± 9.3 pp, low confidence
Fugu Cyber86.9%17.8%estimated ± 9.3 pp, low confidence