ARC-AGI
ARC-AGI · the original Abstraction and Reasoning Corpus, testing whether AI can solve novel visual pattern recognition tasks without memorization.
The Frontier
Best score over time · one chart, every benchmark
Distribution
Where models cluster
Correlated benchmarks
Pearson r · original research
Benchmarks that track with ARC-AGI
Pearson correlation across models scored on both benchmarks. Closer to 1 = strongly predictive.
Full rankings
48 models tested · sorted by score
| # | Model | Score |
|---|---|---|
| 1 |  Gemini 3.1 Pro Preview | 98.0 |
| 2 |  GPT-5.4 Pro | 94.5 |
| 3 |  Claude Opus 4.6 | 94.0 |
| 4 | GPT-5.4 | 93.7 |
| 5 | GPT-5.2 Pro | 90.5 |
| 6 | Claude Sonnet 4.6 | 86.5 |
| 7 | GPT-5.2 | 86.2 |
| 8 | Claude Opus 4.5 | 80.0 |
| 9 | Gemini 3 Pro | 75.0 |
| 10 | GPT-5.1 | 72.8 |
| 11 | GPT-5 Pro | 70.2 |
| 12 |  Grok 4 | 66.7 |
| 13 | GPT-5 | 65.7 |
| 14 |  Kimi K2.5 | 65.3 |
| 15 | Claude Sonnet 4.5 | 63.7 |
| 16 |  MiniMax M2.5 | 63.7 |
| 17 | o3 | 60.8 |
| 18 | o3 Pro | 59.3 |
| 19 | o4 Mini | 58.7 |
| 20 |  DeepSeek V3.2 | 57.0 |
| 21 | GPT-5 Mini | 54.3 |
| 22 | Grok 4 Fast | 48.5 |
| 23 | Claude Haiku 4.5 | 47.7 |
| 24 |  GLM 5 | 44.7 |
| 25 | Gemini 2.5 Pro | 41.0 |
| 26 | Claude Sonnet 4 | 40.0 |
| 27 | Claude Opus 4 | 35.7 |
| 28 | o3 Mini | 34.5 |
| 29 | Gemini 2.5 Flash | 32.3 |
| 30 | o1 | 30.7 |
| 31 | Claude 3.7 Sonnet | 28.6 |
| 32 | Gemini 3 Flash Preview | 21.5 |
| 33 | R1 0528 | 21.2 |
| 34 | GPT-5 Nano | 20.7 |
| 35 | o1-preview | 18.0 |
| 36 | Grok 3 Mini | 16.5 |
| 37 | R1 | 15.8 |
| 38 | o1-mini | 14.0 |
| 39 |  Qwen3 235B A22B Instruct 2507 | 11.0 |
| 40 | GPT-4.5 | 10.3 |
| 41 | GPT-4.1 | 5.5 |
| 42 | Grok 3 | 5.5 |
| 43 | U Magistral Small 1.1 | 5.0 |
| 44 | GPT-4o (2024-11-20) | 4.5 |
| 45 |  Llama 4 Maverick | 4.4 |
| 46 | GPT-4.1 Mini | 3.5 |
| 47 | Llama 4 Scout | 0.5 |
| 48 | GPT-4.1 Nano | 0.1 |
Frequently asked
Pulled from the ARC-AGI dataset · updated daily
What does ARC-AGI measure?
ARC-AGI is a reasoning benchmark in the BenchGecko catalog. 48 AI models have been tested on it. Scores range from 0.1 to 98.0 out of 100.
Which model leads on ARC-AGI?
Gemini 3.1 Pro Preview from Google DeepMind leads ARC-AGI with a score of 98.0. The median score across 48 tested models is 42.8.
Is ARC-AGI saturated?
Yes · the top model on ARC-AGI has reached 98.0 out of 100, within 5% of the theoretical ceiling. This benchmark is approaching saturation and may be replaced by a harder successor.
Does ARC-AGI predict performance on other benchmarks?
Yes · ARC-AGI scores correlate 0.94 with Cybench across 13 shared models. Models that do well on ARC-AGI tend to do well on Cybench.
How often is ARC-AGI data refreshed?
BenchGecko pulls updates daily. New model scores on ARC-AGI appear as soon as they are published by Epoch AI or the model provider.
More reasoning benchmarks
Same category · related evaluations