Calibration
HealthBench Professional → Vals MMLU-Pro
Vals MMLU-Pro is estimated from HealthBench Professional with a Michaelis–Menten curve fitted on 5 models measured on both: y = 2.0000·x / (0.73920 + x), R² = 0.66, cross-validated error 1.5 pp. It is used for 7 estimates.
| Estimated model | HealthBench Professional | Vals MMLU-Pro | Source |
|---|---|---|---|
| Claude Opus 5.5 | 65.6% | 94.0% | estimated ± 1.5 pp, low confidence |
| Claude Sonnet 5.5 | 69.2% | 96.7% | estimated ± 1.5 pp, low confidence |
| GPT-6.1 Sol | 64.2% | 93.0% | estimated ± 1.5 pp, low confidence |
| GPT-6 Luna | 60.8% | 90.3% | estimated ± 1.5 pp, low confidence |
| GPT-6 Sol | 60.8% | 90.3% | estimated ± 1.5 pp, low confidence |
| Grok 4.7 | 56.7% | 86.8% | estimated ± 1.5 pp, medium confidence |
| Ling 3.1 Flash | 65.4% | 93.8% | estimated ± 1.5 pp, low confidence |