benchgap
Calibration

CyberGym → DeepSearchQA

DeepSearchQA is estimated from CyberGym with a offset logistic curve fitted on 5 models measured on both: y = 0.7675 + (1.2000 − 0.7675) / (1 + exp(−199.64·(x − 0.8661))), R² = 0.77, cross-validated error 11.5 pp. It is used for 5 estimates.

Estimated modelCyberGymDeepSearchQASource
Claude Mythos Preview83.1%76.8%estimated ± 11.5 pp, low confidence
Gemini 3.5 Flash Cyber83.2%76.8%estimated ± 11.5 pp, low confidence
Gemini 3.8 Flash Cyber86.2%90.0%estimated ± 11.5 pp, low confidence
GLM-543.2%76.8%estimated ± 11.5 pp, low confidence
Fugu Cyber86.9%100.0%estimated ± 11.5 pp, low confidence