Calibration
HLE → SuperGPQA
SuperGPQA is estimated from HLE with a Michaelis–Menten curve fitted on 12 models measured on both: y = 1.0918·x / (0.15907 + x), R² = 0.72, cross-validated error 9.1 pp. It is used for 11 estimates.
| Estimated model | HLE | SuperGPQA | Source |
|---|---|---|---|
| Agents-A1 | 47.6% | 81.8% | estimated ± 9.1 pp, medium confidence |
| Claude Mythos 5 | 64.5% | 87.6% | estimated ± 9.1 pp, low confidence |
| dots3-note Preview | 52.6% | 83.8% | estimated ± 9.1 pp, medium confidence |
| GPT-5.4 Pro | 58.7% | 85.9% | estimated ± 9.1 pp, low confidence |
| GPT-5.5 Pro | 57.2% | 85.4% | estimated ± 9.1 pp, low confidence |
| Hy4 preview | 55.4% | 84.8% | estimated ± 9.1 pp, low confidence |
| Ornith-1.5-35B-A3B | 25.6% | 67.3% | estimated ± 9.1 pp, medium confidence |
| Ornith-1.5-397B | 44.6% | 80.5% | estimated ± 9.1 pp, medium confidence |
| Ornith-1.5-9B | 20.2% | 61.1% | estimated ± 9.1 pp, medium confidence |
| Qwen3.8 Max | 43.6% | 80.0% | estimated ± 9.1 pp, medium confidence |
| Qwen3.8-Omni-Flash | 36.5% | 76.0% | estimated ± 9.1 pp, medium confidence |