benchgap
Calibration

GPQA Diamond → CritPt

CritPt is estimated from GPQA Diamond with a inverse Michaelis–Menten curve fitted on 16 models measured on both: y = 0.70348·(x − 0.8531) / (0.8531 + 0.3561 − x), R² = 0.71, cross-validated error 6.8 pp. It is used for 3 estimates.

Estimated modelGPQA DiamondCritPtSource
GPT-6 Astra (high)94.9%25.9%estimated ± 6.8 pp, medium confidence
Grok 4.6 (high)94.9%25.9%estimated ± 6.8 pp, medium confidence
Gemini 3.7 Flash (high)94.5%24.5%estimated ± 6.8 pp, medium confidence