Calibration
HealthBench Professional → Vals GPQA Diamond
Vals GPQA Diamond is estimated from HealthBench Professional with a offset logistic curve fitted on 5 models measured on both: y = 0.9112 + (0.9825 − 0.9112) / (1 + exp(−200.00·(x − 0.6034))), R² = 0.91, cross-validated error 1.5 pp. It is used for 7 estimates.
| Estimated model | HealthBench Professional | Vals GPQA Diamond | Source |
|---|---|---|---|
| Claude Opus 5.5 | 65.6% | 98.2% | estimated ± 1.5 pp, low confidence |
| Claude Sonnet 5.5 | 69.2% | 98.2% | estimated ± 1.5 pp, low confidence |
| GPT-6.1 Sol | 64.2% | 98.2% | estimated ± 1.5 pp, low confidence |
| GPT-6 Luna | 60.8% | 96.2% | estimated ± 1.5 pp, low confidence |
| GPT-6 Sol | 60.8% | 96.2% | estimated ± 1.5 pp, low confidence |
| Grok 4.7 | 56.7% | 91.1% | estimated ± 1.5 pp, medium confidence |
| Ling 3.1 Flash | 65.4% | 98.2% | estimated ± 1.5 pp, low confidence |