ARC-AGI-3 (ARC Prize): leaderboard
Third generation of ARC-AGI, released 2026 as part of ARC Prize 2026. Harder abstract reasoning puzzles testing fluid intelligence and generalization from minimal examples.
Metric: Accuracy (%). Source: arcprize.org. Status: years away from saturation. 39 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-6 (Max) | 62.71 |
| 2 | GPT-6 (xHigh) | 59.34 |
| 3 | GPT-6 (High) | 54.82 |
| 4 | GPT-6 (Medium) | 38.59 |
| 5 | GPT-6 (Non-reasoning) | 35.18 |
| 6 | Claude Opus 5 (High) | 30.16 |
| 7 | GPT-6 (Low) | 17.45 |
| 8 | GPT-5.6 Sol (Max) | 7.78 |
| 9 | GPT-5.6 Sol (xHigh) | 6.99 |
| 10 | GPT-5.6 Sol (High) | 2.15 |
| 11 | Grok 4.6 (xHigh) | 2.11 |
| 12 | Claude Opus 4.8 (High) | 1.52 |
| 13 | GPT-5.6 Sol (Medium) | 1.07 |
| 14 | GPT-5.6 Terra (Max) | 0.8 |
| 15 | GPT-5.6 Terra (xHigh) | 0.65 |
Interactive version: theaggregate.ai/benchmark?slug=arc-agi-3-arc-prize · How It Works · Data refreshed daily, snapshot 2026-09-05.