Calibration
HealthBench Professional → GPQA Diamond
GPQA Diamond is estimated from HealthBench Professional with a offset logistic curve fitted on 5 models measured on both: y = 0.9257 + (0.9603 − 0.9257) / (1 + exp(−103.12·(x − 0.6015))), R² = 0.98, cross-validated error 1.0 pp. It is used for 9 estimates.
| Estimated model | HealthBench Professional | GPQA Diamond | Source |
|---|---|---|---|
| Claude Opus 5 | 59.8% | 94.0% | estimated ± 1.0 pp, medium confidence |
| Claude Opus 5.5 | 65.6% | 96.0% | estimated ± 1.0 pp, low confidence |
| Claude Sonnet 5.5 | 69.2% | 96.0% | estimated ± 1.0 pp, low confidence |
| GPT-6.1 Sol | 64.2% | 96.0% | estimated ± 1.0 pp, medium confidence |
| GPT-6 Luna | 60.8% | 94.9% | estimated ± 1.0 pp, medium confidence |
| GPT-6 Sol | 60.8% | 94.9% | estimated ± 1.0 pp, medium confidence |
| Grok 4.7 | 56.7% | 92.7% | estimated ± 1.0 pp, medium confidence |
| Ling 3.1 Flash | 65.4% | 96.0% | estimated ± 1.0 pp, low confidence |
| Muse Spark 1.1 | 59.3% | 93.6% | estimated ± 1.0 pp, medium confidence |