Calibration
ARC-AGI-2 → GPQA
GPQA is estimated from ARC-AGI-2 with a Hill curve fitted on 18 models measured on both: y = 0.8062 + (1.2000 − 0.8062)·x^0.79 / (2.02018^0.79 + x^0.79), R² = 0.84, cross-validated error 1.6 pp. It is used for 8 estimates.
| Estimated model | ARC-AGI-2 | GPQA | Source |
|---|---|---|---|
| Claude Opus 4.5 Thinking | 37.6% | 88.9% | estimated ± 1.6 pp, high confidence |
| Claude Opus 4.6 (Adaptive) | 68.8% | 92.4% | estimated ± 1.6 pp, high confidence |
| Claude Sonnet 4.5 Thinking | 13.6% | 84.8% | estimated ± 1.6 pp, high confidence |
| dots3-note Preview | 81.4% | 93.6% | estimated ± 1.6 pp, high confidence |
| Gemini 3 Pro | 31.1% | 88.0% | estimated ± 1.6 pp, high confidence |
| Gemini 3 Pro Deep Think | 45.1% | 89.9% | estimated ± 1.6 pp, high confidence |
| GPT-5.4 Pro | 83.3% | 93.7% | estimated ± 1.6 pp, high confidence |
| GPT-5.5 Pro | 84.2% | 93.8% | estimated ± 1.6 pp, high confidence |