Calibration
ARC-AGI-3 → HLE w/o tools
HLE w/o tools is estimated from ARC-AGI-3 with a Michaelis–Menten curve fitted on 7 models measured on both: y = 0.5275·x / (0.00056 + x), R² = 0.75, cross-validated error 5.0 pp. It is used for 6 estimates.
| Estimated model | ARC-AGI-3 | HLE w/o tools | Source |
|---|---|---|---|
| Gemini 3.8 Flash | 10.4% | 52.5% | estimated ± 5.0 pp, medium confidence |
| GPT-6.1 Sol | 52.7% | 52.7% | estimated ± 5.0 pp, low confidence |
| GPT-6 Luna | 0.1% | 33.8% | estimated ± 5.0 pp, medium confidence |
| GPT-6 Sol | 4.6% | 52.1% | estimated ± 5.0 pp, medium confidence |
| Grok 4.5 | 0.3% | 44.5% | estimated ± 5.0 pp, medium confidence |
| Grok 4.6 | 2.1% | 51.4% | estimated ± 5.0 pp, medium confidence |