ARC-AGI-1: leaderboard
800 grid puzzles (400 train, 400 eval) from François Chollet (2019): infer a rule from about three input-output grids and draw the output for a new input; o3-preview scored 75-87% in Dec 2024.
Metric: Accuracy (%). Source: arcprize.org. Status: saturated. 223 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Claude Fable 5 (Max) | 98.5 |
| 2 | GPT-6 (xHigh) | 98.5 |
| 3 | GPT-6 (High) | 98.5 |
| 4 | Gemini 3.1 Pro (Preview) | 98 |
| 5 | Claude Opus 5 (Max) | 97.5 |
| 6 | GPT-5.6 Sol (xHigh) | 97.5 |
| 7 | Claude Fable 5.1 (Max) | 97.5 |
| 8 | Claude Opus 5 (High) | 97.5 |
| 9 | GPT-6 (Max) | 97.5 |
| 10 | GPT-6 (Medium) | 97.5 |
| 11 | GPT-5.6 Sol (High) | 97 |
| 12 | GPT-5.6 Sol (Max) | 96.5 |
| 13 | GPT-5.6 Terra (Max) | 96.5 |
| 14 | GPT-6 (Low) | 96.5 |
| 15 | Claude Fable 5.1 (xHigh) | 96.5 |
Interactive version: theaggregate.ai/benchmark?slug=arc-agi-1 · How It Works · Data refreshed daily, snapshot 2026-09-05.