Calibration
SuperGPQA → MMLU-Pro
MMLU-Pro is estimated from SuperGPQA with a Michaelis–Menten curve fitted on 19 models measured on both: y = 1.1467·x / (0.24027 + x), R² = 0.76, cross-validated error 6.3 pp. It is used for 2 estimates.
| Estimated model | SuperGPQA | MMLU-Pro | Source |
|---|---|---|---|
| Seed 2.1 Pro | 70.8% | 85.6% | estimated ± 6.3 pp, medium confidence |
| Seed 2.1 Turbo | 67.4% | 84.5% | estimated ± 6.3 pp, medium confidence |
Fitted on
| Model | SuperGPQA | MMLU-Pro |
|---|---|---|
| Claude Opus 4.5 | 70.6% | 89.5% |
| Claude Opus 4.6 | 95.0% | 82.0% |
| Claude Sonnet 4.6 | 95.0% | 79.2% |
| GLM-5 | 66.8% | 85.7% |
| Kimi K2 | 57.2% | 81.1% |
| Kimi K2.5 | 69.2% | 87.1% |
| MiniCPM5-1B | 23.1% | 48.9% |
| MiniCPM5-2B | 40.8% | 70.8% |
| Qwen3 235B 2507 | 62.6% | 83.0% |
| Qwen3 235B 2507 (Reasoning) | 64.9% | 84.4% |
| Qwen3.5-122B-A10B | 67.1% | 86.7% |
| Qwen3.5-27B | 65.6% | 86.1% |
| Qwen3.5-35B-A3B | 63.4% | 85.3% |
| Qwen3.5 397B | 70.4% | 87.8% |
| Qwen3.6-27B | 66.0% | 86.2% |
| Qwen3.6-35B-A3B | 64.7% | 85.2% |
| Qwen3.6 Plus | 71.6% | 88.5% |
| Qwen3.7 Max | 73.6% | 89.6% |
| Qwen3.7 Plus | 71.4% | 88.5% |