Calibration
Vals MMLU-Pro → ARC-AGI-3
ARC-AGI-3 is estimated from Vals MMLU-Pro with a inverse Michaelis–Menten curve fitted on 12 models measured on both: y = 0.00041·(x − 0.0000) / (0.0000 + 0.9173 − x), R² = 0.84, cross-validated error 9.0 pp. It is used for 5 estimates.
| Estimated model | Vals MMLU-Pro | ARC-AGI-3 | Source |
|---|---|---|---|
| Gemini 3.1 Flash-Lite | 86.2% | 0.6% | estimated ± 9.0 pp, medium confidence |
| GLM-5.3-Flash | 86.1% | 0.6% | estimated ± 9.0 pp, medium confidence |
| Qwen3.5 Flash | 84.1% | 0.5% | estimated ± 9.0 pp, low confidence |
| MiMo-V2.5 | 82.9% | 0.4% | estimated ± 9.0 pp, low confidence |
| GLM-4.5 | 81.2% | 0.3% | estimated ± 9.0 pp, low confidence |
Fitted on
| Model | Vals MMLU-Pro | ARC-AGI-3 |
|---|---|---|
| Claude Opus 4.7 | 89.9% | 0.2% |
| Claude Opus 4.8 | 89.6% | 1.5% |
| Claude Opus 5 | 91.6% | 30.2% |
| Gemini 3.1 Pro | 91.0% | 0.4% |
| Gemini 3.8 Flash | 90.2% | 10.4% |
| GPT-5.5 | 88.1% | 0.4% |
| GPT-5.6 Luna | 86.0% | 0.2% |
| GPT-5.6 Sol | 89.1% | 7.8% |
| GPT-5.6 Terra | 86.7% | 0.8% |
| Grok 4.20 | 86.3% | 0.1% |
| Grok 4.5 | 89.2% | 0.3% |
| Grok 4.6 | 89.4% | 2.1% |