Calibration
GPQA → MMLU-Redux
MMLU-Redux is estimated from GPQA with a Michaelis–Menten curve fitted on 9 models measured on both: y = 1.1388·x / (0.18732 + x), R² = 0.90, cross-validated error 2.1 pp. It is used for 7 estimates.
| Estimated model | GPQA | MMLU-Redux | Source |
|---|---|---|---|
| Claude 3.5 Sonnet | 59.4% | 86.6% | estimated ± 2.1 pp, high confidence |
| Claude Sonnet 4.5 | 83.4% | 93.0% | estimated ± 2.1 pp, high confidence |
| Interfaze Beta | 89.9% | 94.2% | estimated ± 2.1 pp, high confidence |
| o1-pro | 79.0% | 92.1% | estimated ± 2.1 pp, high confidence |
| o3-mini | 77.2% | 91.6% | estimated ± 2.1 pp, high confidence |
| Sakana Fugu | 95.5% | 95.2% | estimated ± 2.1 pp, medium confidence |
| Sakana Fugu-Ultra | 95.5% | 95.2% | estimated ± 2.1 pp, medium confidence |