Calibration
Vals MMLU-Pro → ARC-AGI-1
ARC-AGI-1 is estimated from Vals MMLU-Pro with a linear curve fitted on 21 models measured on both: y = 3.0952·x + -1.8337, R² = 0.76, cross-validated error 6.3 pp. It is used for 4 estimates.
| Estimated model | Vals MMLU-Pro | ARC-AGI-1 | Source |
|---|---|---|---|
| Muse Spark 1.2 | 88.3% | 89.9% | estimated ± 6.3 pp, medium confidence |
| Mistral Medium 3.5 128B | 75.3% | 49.7% | estimated ± 6.3 pp, low confidence |
| Laguna XS.2 | 69.1% | 30.5% | estimated ± 6.3 pp, low confidence |
| Laguna M.1 | 68.8% | 29.6% | estimated ± 6.3 pp, low confidence |
Fitted on
| Model | Vals MMLU-Pro | ARC-AGI-1 |
|---|---|---|
| Claude Fable 5 | 91.5% | 98.5% |
| Claude Fable 5.1 | 92.4% | 97.5% |
| Claude Opus 4.8 | 89.6% | 92.5% |
| Claude Opus 5 | 91.6% | 97.5% |
| Claude Sonnet 4.6 | 87.3% | 86.0% |
| DeepSeek V4 Flash 0731 | 86.2% | 89.0% |
| DeepSeek V4 Pro 0813 | 87.0% | 90.0% |
| Gemini 3.5 Flash | 89.5% | 92.5% |
| Gemini 3.6 Flash | 89.3% | 91.2% |
| Gemini 3.7 Flash | 90.1% | 95.5% |
| Gemini 3.8 Flash | 90.2% | 98.5% |
| GPT-5.4 mini | 84.6% | 63.7% |
| GPT-5.4 nano | 77.2% | 51.5% |
| GPT-5.5 | 88.1% | 95.0% |
| GPT-5.6 Luna | 86.0% | 88.0% |
| GPT-5.6 Sol | 89.1% | 96.5% |
| GPT-5.6 Terra | 86.7% | 96.5% |
| Grok 4.5 | 89.2% | 85.7% |
| Grok 4.6 | 89.4% | 87.0% |
| Inkling-Small | 85.6% | 84.0% |
| Kimi K3 | 88.0% | 94.5% |