ARC-AGI-1: leaderboard
Grid puzzles in which a model sees a small number of input and output grid pairs, infers the transformation, and then draws the output grid for a new input. A task counts as solved only when every cell of the produced grid matches. The ARC Prize leaderboard runs models on the 100-task semi-private evaluation set with a single run rather than an average, and reports the share of tasks solved, higher being better.
Metric: Accuracy (%). Source: arcprize.org. Source: arcprize.org. Saturation forecast: Estimated already saturated. 275 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.8 Flash (High) | 98.5 | #21 (Gemini 3.8 Flash) |
| 2 | GPT-6 Astra (xHigh) | 98.5 | #6 (GPT-6 Astra) |
| 3 | GPT-6 Astra (High) | 98.5 | #6 (GPT-6 Astra) |
| 4 | Gemini 4 Argon (High) | 98.5 | #8 (Gemini 4 Argon) |
| 5 | Claude Fable 5 (Max) | 98.5 | #16 (Claude Fable 5) |
| 6 | GPT-6.1 Sol (xHigh) | 98.5 | #7 (GPT-6.1 Sol) |
| 7 | GPT-6.1 Sol (High) | 98.5 | #7 (GPT-6.1 Sol) |
| 8 | Claude Opus 5.5 (High) | 98.5 | #3 (Claude Opus 5.5) |
| 9 | Claude Fable 5 (xHigh) | 98.5 | #16 (Claude Fable 5) |
| 10 | Gemini 4 Argon (Medium) | 98.5 | #8 (Gemini 4 Argon) |
| 11 | Gemini 3.1 Pro (Preview) | 98 | #58 |
| 12 | Claude Opus 5 (Max) | 97.5 | #12 (Claude Opus 5) |
| 13 | GPT-6 Astra (Max) | 97.5 | #6 (GPT-6 Astra) |
| 14 | GPT-5.6 Sol (xHigh) | 97.5 | #19 (GPT-5.6 Sol) |
| 15 | Claude Opus 5 (High) | 97.5 | #12 (Claude Opus 5) |
No result here: #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=arc-agi-1 · How It Works · Data refreshed daily, snapshot 2026-10-20.