// benchmark explainer
ARC-AGI, explained
By the TrueIQ editorial team at Ahzy Capital · Last reviewed Oct 15, 2026 · How we work
ARC-AGI shows a model a few example grids that follow a hidden rule and asks it to apply the rule to a new grid. Most people find them easy, but they have been hard for AI.
| Created by | François Chollet (2019); run by the ARC Prize Foundation |
|---|---|
| Format | Coloured grid input/output puzzles |
| Versions | ARC-AGI-1, ARC-AGI-2 and later |
| Prize | ARC Prize competitions with private test sets |
What it measures
It aims to measure skill acquisition efficiency: learning something genuinely new from very few examples instead of recalling training data.
Where it falls short
- The grid format is narrow and quite different from everyday tasks.
- Reported results often depend on large amounts of computation per task. ARC Prize reports cost alongside scores for that reason.
- Public versus semi private versus private sets give different numbers.
How to read a score
Check which version and which test set (public or private) a score uses, and how much compute was spent per task.
Where to see current scores
We don't republish scores. Leaderboards change weekly, and numbers depend on how the test was run. Check these instead: