Calibration
ARC-AGI-2 → HLE
HLE is estimated from ARC-AGI-2 with a Hill curve fitted on 18 models measured on both: y = 0.4384 + (1.2000 − 0.4384)·x^6.00 / (1.08577^6.00 + x^6.00), R² = 0.56, cross-validated error 6.3 pp. It is used for 10 estimates.
| Estimated model | ARC-AGI-2 | HLE | Source |
|---|---|---|---|
| Claude Fable 5 | 89.2% | 61.7% | estimated ± 6.3 pp, medium confidence |
| Claude Opus 4.5 Thinking | 37.6% | 44.0% | estimated ± 6.3 pp, medium confidence |
| Claude Opus 4.6 (Adaptive) | 68.8% | 48.5% | estimated ± 6.3 pp, medium confidence |
| Claude Sonnet 4.5 | 13.6% | 43.8% | estimated ± 6.3 pp, medium confidence |
| Claude Sonnet 4.5 Thinking | 13.6% | 43.8% | estimated ± 6.3 pp, medium confidence |
| Gemini 3.6 Flash | 60.4% | 46.0% | estimated ± 6.3 pp, medium confidence |
| Gemini 3.7 Flash | 84.6% | 57.8% | estimated ± 6.3 pp, medium confidence |
| Gemini 3 Pro | 31.1% | 43.9% | estimated ± 6.3 pp, medium confidence |
| Gemini 3 Pro Deep Think | 45.1% | 44.2% | estimated ± 6.3 pp, medium confidence |
| GPT-5.2 | 52.9% | 44.8% | estimated ± 6.3 pp, medium confidence |