benchgap
Calibration

CritPt → AA-HLE

AA-HLE is estimated from CritPt with a Hill curve fitted on 186 models measured on both: y = 0.0733 + (1.2000 − 0.0733)·x^0.53 / (0.71180^0.53 + x^0.53), R² = 0.89, cross-validated error 4.6 pp. It is used for 3 estimates.

Estimated modelCritPtAA-HLESource
Gemini 3 Pro Deep Think25.7%48.8%estimated ± 4.6 pp, high confidence
GPT-5.4 Pro30.0%51.0%estimated ± 4.6 pp, high confidence
GPT-5.5 Pro30.6%51.3%estimated ± 4.6 pp, high confidence