Calibration
GPQA → SuperGPQA
SuperGPQA is estimated from GPQA with a Hill curve fitted on 15 models measured on both: y = 0.2922 + (1.2000 − 0.2922)·x^6.00 / (0.89406^6.00 + x^6.00), R² = 0.35, cross-validated error 9.2 pp. It is used for 13 estimates.
| Estimated model | GPQA | SuperGPQA | Source |
|---|---|---|---|
| Claude 3.5 Sonnet | 59.4% | 36.4% | estimated ± 9.2 pp, low confidence |
| Claude Sonnet 4.5 | 83.4% | 65.3% | estimated ± 9.2 pp, low confidence |
| Interfaze Beta | 89.9% | 75.4% | estimated ± 9.2 pp, low confidence |
| LFM2.5-230M | 25.4% | 29.3% | estimated ± 9.2 pp, low confidence |
| LFM2.5-VL-450M | 25.7% | 29.3% | estimated ± 9.2 pp, low confidence |
| MAI-Thinking-1 | 84.2% | 66.5% | estimated ± 9.2 pp, low confidence |
| o1-pro | 79.0% | 58.5% | estimated ± 9.2 pp, low confidence |
| o3-mini | 77.2% | 55.8% | estimated ± 9.2 pp, low confidence |
| Sakana Fugu | 95.5% | 83.5% | estimated ± 9.2 pp, low confidence |
| Sakana Fugu-Ultra | 95.5% | 83.5% | estimated ± 9.2 pp, low confidence |
| Soofi S 30B-A3B | 43.4% | 30.4% | estimated ± 9.2 pp, low confidence |
| ZAYA1-74B-Preview | 57.3% | 35.1% | estimated ± 9.2 pp, low confidence |
| ZAYA1-8B | 71.0% | 47.4% | estimated ± 9.2 pp, low confidence |