ARC-AGI-2: leaderboard
Each task shows a handful of example grids with their transformed outputs and asks the model to infer the unstated rule and produce the output grid for a new input, on grids of up to 30 by 30 cells in ten colours. The board reports the solve rate with two attempts per task on the 120-task semi-private set; the source leaderboard also lists each run's cost per task. Higher is better. Source: arcprize.org/arc-agi/2.
Metric: Accuracy (%). Source: arcprize.org. Saturation forecast: Around January 2027. 277 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Best Expert | 100 | |
| 2 | GPT-6 Astra (Max) | 95 | #6 (GPT-6 Astra) |
| 3 | GPT-6.1 Sol (Max) | 94.17 | #7 (GPT-6.1 Sol) |
| 4 | GPT-6 Astra (xHigh) | 93.33 | #6 (GPT-6 Astra) |
| 5 | Claude Opus 5.5 (High) | 93.33 | #3 (Claude Opus 5.5) |
| 6 | GPT-5.6 Sol (Max) | 92.5 | #19 (GPT-5.6 Sol) |
| 7 | Claude Opus 5.5 (xHigh) | 92.5 | #3 (Claude Opus 5.5) |
| 8 | GPT-6 Astra (High) | 92.08 | #6 (GPT-6 Astra) |
| 9 | GPT-6 Astra (Medium) | 92.08 | #6 (GPT-6 Astra) |
| 10 | GPT-6.1 Sol (xHigh) | 91.67 | #7 (GPT-6.1 Sol) |
| 11 | Claude Opus 5.5 (Max) | 91.67 | #3 (Claude Opus 5.5) |
| 12 | GPT-6.1 Sol (High) | 91.67 | #7 (GPT-6.1 Sol) |
| 13 | Gemini 4 Argon (High) | 91.25 | #8 (Gemini 4 Argon) |
| 14 | Claude Opus 5 (Max) | 90.42 | #12 (Claude Opus 5) |
| 15 | GPT-5.6 Sol (xHigh) | 90 | #19 (GPT-5.6 Sol) |
No result here: #10 Muse Wonder Full.
Interactive version: theaggregate.ai/benchmark?slug=arc-agi-2 · How It Works · Data refreshed daily, snapshot 2026-10-20.