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
| Model | HealthBench (raw) | AA-HLE |
|---|---|---|
| Claude Opus 5 | 67.1% | 54.9% |
| Claude Opus 5.5 | 68.1% | 61.4% |
| Claude Sonnet 5.5 | 69.4% | 55.0% |
| GPT-6.1 Sol | 56.7% | 52.9% |
| GPT-6.1 Terra | 51.1% | 46.2% |
| GPT-6 Astra | 56.9% | 54.7% |
| GPT-6 Luna | 50.0% | 38.5% |
| GPT-6 Sol | 47.1% | 47.9% |