benchgap
Calibration

HealthBench (raw) → AA-HLE

AA-HLE is estimated from HealthBench (raw) with a Michaelis–Menten curve fitted on 8 models measured on both: y = 1.7872·x / (1.43411 + x), R² = 0.64, cross-validated error 5.1 pp. It is used for 0 estimates.

Fitted on

ModelHealthBench (raw)AA-HLE
Claude Opus 567.1%54.9%
Claude Opus 5.568.1%61.4%
Claude Sonnet 5.569.4%55.0%
GPT-6.1 Sol56.7%52.9%
GPT-6.1 Terra51.1%46.2%
GPT-6 Astra56.9%54.7%
GPT-6 Luna50.0%38.5%
GPT-6 Sol47.1%47.9%