- Cleus
- AI benchmarks
- ARC-AGI-2
ARC-AGI-2explained
Puzzles that are easy for people and hard for AI. Each ARC-AGI-2 task shows a few small coloured grids and what they turn into. The model has to work out the hidden rule and apply it to a new grid. Every task was solved by real people in testing, usually in a couple of tries.
- Highest published score
- GPT-6 Astra, 95.0%
- Leader among popular models
- GPT-6 Astra
- Models with a result
- 94
- Skill it tests
- Reasoning
What ARC-AGI-2 really tests
- What it is
- Each ARC-AGI-2 task shows a few small coloured grids and what they turn into. The model has to work out the hidden rule and apply it to a new grid. Every task was solved by real people in testing, usually in a couple of tries.
- A question like
- In the examples, every red shape gets a blue outline and shapes touching the edge disappear. Given a new grid, produce the right output, cell for cell. Our own example in the style of the test, not a real item.
- Why it matters
- It tests learning a brand-new rule from a handful of examples, which is close to what people mean by “general intelligence”. Memorising the internet doesn’t help here.
- Reading the score
- Share of puzzles solved exactly. Every task was solved by at least two people in testing, and panels of people solve almost all of them.
- The catch
- Scores depend a lot on how much computing time the model is allowed per puzzle, and the ARC Prize team publishes cost per task alongside scores. A high score at a huge cost is less impressive than it looks.
- Made by
- ARC Prize Foundation (François Chollet), 2025. Official page
Leaderboard
Each model’s best published result across its reasoning settings.
Popular models
The models we track across this site.
Top 20 of all 94 models tested
Includes older, preview and specialist models that aren’t on our main leaderboard.
Questions
What is ARC-AGI-2?
Each ARC-AGI-2 task shows a few small coloured grids and what they turn into. The model has to work out the hidden rule and apply it to a new grid. Every task was solved by real people in testing, usually in a couple of tries.
What does a ARC-AGI-2 score mean?
Share of puzzles solved exactly. Every task was solved by at least two people in testing, and panels of people solve almost all of them.
Which AI model is best on ARC-AGI-2?
GPT-6 Astra from OpenAI has the highest published ARC-AGI-2 result we have (95.0%).
Is ARC-AGI-2 a good benchmark?
It tests learning a brand-new rule from a handful of examples, which is close to what people mean by “general intelligence”. Memorising the internet doesn’t help here. The catch: Scores depend a lot on how much computing time the model is allowed per puzzle, and the ARC Prize team publishes cost per task alongside scores. A high score at a huge cost is less impressive than it looks.
Who made ARC-AGI-2?
ARC Prize Foundation (François Chollet), 2025.
Benchmark scores: Epoch AI, Capabilities & benchmarking, used under CC BY 4.0. For each model we show its best published result across reasoning settings. Prices and context sizes: OpenRouter. Data refreshed November 9, 2026. Hands-on tests (answers, websites, code, refusals, speed) are our own single runs. Explanations written by Cleus.