Calibration
SuperGPQA → C-Eval
C-Eval is estimated from SuperGPQA with a Hill curve fitted on 6 models measured on both: y = 0.0000 + (0.9732 − 0.0000)·x^6.00 / (0.42105^6.00 + x^6.00), R² = 0.71, cross-validated error 1.0 pp. It is used for 4 estimates.
| Estimated model | SuperGPQA | C-Eval | Source |
|---|---|---|---|
| Seed 2.1 Pro | 70.8% | 93.2% | estimated ± 1.0 pp, medium confidence |
| Seed 2.1 Turbo | 67.4% | 91.9% | estimated ± 1.0 pp, medium confidence |
| Qwen3.5-122B-A10B | 67.1% | 91.7% | estimated ± 1.0 pp, medium confidence |
| Qwen3 235B 2507 | 62.6% | 89.1% | estimated ± 1.0 pp, low confidence |
Fitted on
| Model | SuperGPQA | C-Eval |
|---|---|---|
| Claude Opus 4.5 | 70.6% | 92.2% |
| Kimi K2.5 | 69.2% | 94.0% |
| Qwen3.5 397B | 70.4% | 93.0% |
| Qwen3.6-27B | 66.0% | 91.4% |
| Qwen3.6-35B-A3B | 64.7% | 90.0% |
| Qwen3.6 Plus | 71.6% | 93.3% |