Calibration
HealthBench (length-adjusted) → ARC-AGI-3
ARC-AGI-3 is estimated from HealthBench (length-adjusted) with a linear curve fitted on 5 models measured on both: y = 10.4190·x + -5.5819, R² = 0.84, cross-validated error 14.4 pp. It is used for 0 estimates.
Fitted on
| Model | HealthBench (length-adjusted) | ARC-AGI-3 |
|---|---|---|
| Claude Opus 5 | 57.8% | 30.2% |
| GPT-6.1 Sol | 58.5% | 52.7% |
| GPT-6 Astra | 58.3% | 62.7% |
| GPT-6 Luna | 54.5% | 0.1% |
| GPT-6 Sol | 53.2% | 4.6% |