Skip to content

ARC-AGI

A dataset of abstraction and reasoning tasks for AI

Visit official site
Category AI models

What ARC-AGI is

The dataset is described as a way to test abstraction and reasoning in AI systems. It contains 400 training tasks and 400 evaluation tasks stored as JSON pairs.

What you can do with it

  • Train a model to solve abstract reasoning tasks
  • Use the task set as a teaching tool about AI
  • Benchmark an AI product on the evaluation tasks

Getting started

  • Download the dataset from the GitHub page
  • Pick a task JSON file and read its input-output pairs
  • Open the test interface in a browser and load the task

What to watch out for

  • Each test output grid allows three attempts
  • Grid cells are visualized with integers from 0 to 9

Pros & cons

✓ What we like

  • Three attempts are allowed per output grid
  • Grid cells use integers 0 to 9 for colors

! What to watch out for

  • Tasks are stored as JSON files to be loaded
  • Only three attempts are allowed per test output grid

FAQ

How many tasks are there?

400 training tasks and 400 evaluation tasks.

What format are tasks in?

JSON files with train and test input-output pairs.

Last reviewed: 2026-11-15

More LLM API platform tools

View all →

How we review