ARC-AGI-2 (ARC Prize): leaderboard
Significantly harder successor to ARC-AGI-1, released 2025. Unseen grid-based puzzles requiring fluid reasoning and generalization from few examples. Frontier models scored under 5% at release; scores have risen steeply since.
Metric: Accuracy (%). Source: arcprize.org. Status: saturation imminent. 225 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-6 (Max) | 95 |
| 2 | GPT-6 (xHigh) | 93.33 |
| 3 | GPT-5.6 Sol (Max) | 92.5 |
| 4 | GPT-6 (Medium) | 92.08 |
| 5 | GPT-6 (High) | 92.08 |
| 6 | Claude Opus 5 (Max) | 90.42 |
| 7 | GPT-5.6 Sol (xHigh) | 90 |
| 8 | Claude Fable 5.1 (Max) | 90 |
| 9 | Claude Fable 5.1 (xHigh) | 90 |
| 10 | Claude Fable 5 (Max) | 89.17 |
| 11 | Claude Fable 5.1 (High) | 88.75 |
| 12 | Claude Opus 5 (High) | 88.33 |
| 13 | Claude Fable 5 (High) | 87.5 |
| 14 | GPT-5.6 Sol (High) | 85.42 |
| 15 | GPT-6 (Low) | 85.42 |
Interactive version: theaggregate.ai/benchmark?slug=arc-agi-2-arc-prize · How It Works · Data refreshed daily, snapshot 2026-09-05.