Calibration
HealthBench Hard → GPQA
GPQA is estimated from HealthBench Hard with a linear curve fitted on 6 models measured on both: y = 0.1148·x + 0.8970, R² = 0.35, cross-validated error 1.7 pp. It is used for 0 estimates.
Fitted on
| Model | HealthBench Hard | GPQA |
|---|---|---|
| Claude Opus 4.6 | 14.8% | 91.3% |
| GPT-5.4 | 40.1% | 92.8% |
| GPT-5.6 Luna | 32.0% | 92.3% |
| GPT-5.6 Sol | 33.1% | 94.6% |
| GPT-5.6 Terra | 32.7% | 92.9% |
| GPT-6 Astra | 36.6% | 96.0% |