Calibration
ARC-AGI-2 → ARC-AGI-1
ARC-AGI-1 is estimated from ARC-AGI-2 with a Michaelis–Menten + offset curve fitted on 34 models measured on both: y = 0.4271 + 0.8169·x / (0.45739 + x), R² = 0.97, cross-validated error 1.9 pp. It is used for 7 estimates.
| Estimated model | ARC-AGI-2 | ARC-AGI-1 | Source |
|---|---|---|---|
| Claude Opus 4.7 (Adaptive) | 75.8% | 93.7% | estimated ± 1.9 pp, high confidence |
| Claude Sonnet 4.5 | 13.6% | 61.4% | estimated ± 1.9 pp, high confidence |
| dots3-note Preview | 81.4% | 95.0% | estimated ± 1.9 pp, high confidence |
| Gemini 3.1 Pro | 77.1% | 94.0% | estimated ± 1.9 pp, high confidence |
| Gemini 3 Pro Deep Think | 45.1% | 83.3% | estimated ± 1.9 pp, high confidence |
| Grok 4.20 | 53.3% | 86.7% | estimated ± 1.9 pp, high confidence |
| Muse Spark | 42.5% | 82.1% | estimated ± 1.9 pp, high confidence |