Calibration
SuperGPQA → GPQA
GPQA is estimated from SuperGPQA with a Hill curve fitted on 17 models measured on both: y = 0.0000 + (0.9413 − 0.0000)·x^6.00 / (0.45275^6.00 + x^6.00), R² = 0.78, cross-validated error 2.5 pp. It is used for 3 estimates.
| Estimated model | SuperGPQA | GPQA | Source |
|---|---|---|---|
| Qwen 3.6 Max (preview) | 73.9% | 89.4% | estimated ± 2.5 pp, high confidence |
| Seed 2.1 Pro | 70.8% | 88.1% | estimated ± 2.5 pp, high confidence |
| Seed 2.1 Turbo | 67.4% | 86.2% | estimated ± 2.5 pp, high confidence |
Fitted on
| Model | SuperGPQA | GPQA |
|---|---|---|
| Claude Opus 4.5 | 70.6% | 87.0% |
| Claude Opus 4.6 | 95.0% | 91.3% |
| Claude Sonnet 4.6 | 95.0% | 89.9% |
| GLM-5 | 66.8% | 86.0% |
| Kimi K2 | 57.2% | 75.1% |
| Kimi K2.5 | 69.2% | 87.6% |
| Qwen3 235B 2507 | 62.6% | 77.5% |
| Qwen3 235B 2507 (Reasoning) | 64.9% | 81.1% |
| Qwen3.5-122B-A10B | 67.1% | 86.6% |
| Qwen3.5-27B | 65.6% | 85.5% |
| Qwen3.5-35B-A3B | 63.4% | 84.2% |
| Qwen3.5 397B | 70.4% | 88.4% |
| Qwen3.6-27B | 66.0% | 87.8% |
| Qwen3.6-35B-A3B | 64.7% | 86.0% |
| Qwen3.6 Plus | 71.6% | 90.4% |
| Qwen3.7 Max | 73.6% | 92.4% |
| Qwen3.7 Plus | 71.4% | 90.3% |