Calibration
FrontierMath v2 (Tier 4) → FrontierMath (legacy)
FrontierMath (legacy) is estimated from FrontierMath v2 (Tier 4) with a Hill curve fitted on 6 models measured on both: y = 0.4466 + (0.9159 − 0.4466)·x^6.00 / (0.50536^6.00 + x^6.00), R² = 1.00, cross-validated error 2.3 pp. It is used for 1 estimate.
| Estimated model | FrontierMath v2 (Tier 4) | FrontierMath (legacy) | Source |
|---|---|---|---|
| GPT-6 Astra | 97.6% | 90.7% | estimated ± 2.3 pp, low confidence |