Calibration
ARC-AGI-3 → GPQA
GPQA is estimated from ARC-AGI-3 with a Michaelis–Menten + offset curve fitted on 8 models measured on both: y = 0.9306 + 0.0340·x / (0.09766 + x), R² = 0.77, cross-validated error 0.8 pp. It is used for 7 estimates.
| Estimated model | ARC-AGI-3 | GPQA | Source |
|---|---|---|---|
| Claude Opus 5 | 30.2% | 95.6% | estimated ± 0.8 pp, high confidence |
| Gemini 3.8 Flash | 10.4% | 94.8% | estimated ± 0.8 pp, high confidence |
| GPT-6.1 Sol | 52.7% | 95.9% | estimated ± 0.8 pp, high confidence |
| GPT-6 Luna | 0.1% | 93.1% | estimated ± 0.8 pp, medium confidence |
| GPT-6 Sol | 4.6% | 94.2% | estimated ± 0.8 pp, high confidence |
| Grok 4.5 | 0.3% | 93.2% | estimated ± 0.8 pp, high confidence |
| Grok 4.6 | 2.1% | 93.7% | estimated ± 0.8 pp, high confidence |