Calibration
PhysicianBench → HealthBench (raw, Artificial Analysis)
HealthBench (raw, Artificial Analysis) is estimated from PhysicianBench with a Michaelis–Menten + offset curve fitted on 5 models measured on both: y = 0.0000 + 0.8920·x / (0.26551 + x), R² = 0.47, cross-validated error 4.3 pp. It is used for 0 estimates.
Fitted on
| Model | PhysicianBench | HealthBench (raw, Artificial Analysis) |
|---|---|---|
| Claude Fable 5.1 | 61.0% | 63.9% |
| Claude Fable 5.2 | 64.4% | 58.8% |
| Claude Opus 5.5 | 68.4% | 64.3% |
| Claude Sonnet 5.5 | 63.2% | 65.9% |
| GPT-6 Astra | 48.0% | 56.9% |