ARC-AGI
In short
ARC-AGI is a set of visual puzzle benchmarks that test whether an AI can work out a new rule from a few examples. The puzzles are easy for most people but have long been hard for AI, which makes it a popular test of general reasoning.
- Also called
- ARC, Abstraction and Reasoning Corpus, ARC Prize, ARC-AGI-2, ARC-AGI-3
- Topic
- Testing and benchmarks
- Level
- Some background helps
- First described
- François Chollet introduced the Abstraction and Reasoning Corpus in "On the Measure of Intelligence", November 20191.
ARC-AGI was created by the Google engineer François Chollet in 2019, in a paper arguing that intelligence means how efficiently a system learns new skills, not how many skills it already has1. The test was first called the Abstraction and Reasoning Corpus and later renamed ARC-AGI2.
Each task shows a few pairs of coloured grids, up to 30 by 30 cells. The output grid follows from the input by some hidden rule. The solver must work out the rule and produce the output for a new input, with two attempts allowed2. Tasks rely only on basic "core knowledge", such as objects and counting, not language or facts3.
The ARC Prize competition launched in 2024 with a target of 85%. During that year the best score on the private test set rose from 33% to 55.5%2. ARC-AGI-2 followed in 2025 with harder tasks; every task was solved by at least two people in two attempts or fewer, in testing with over 400 members of the public45. It also reports cost per task, because brute-force search is not the same as intelligence5.
ARC-AGI-3, launched in March 2026, is interactive. Instead of static puzzles, AI agents must explore small game-like environments and learn goals without written instructions6. The 2026 competition offers $2 million in prizes, with results due on 4 December 20267.
A high ARC score does not prove a system is generally intelligent. Its designers frame it as measuring the gap between human and AI learning6.
Example
A task might show three grids where every blue square has been moved to the bottom row. A person spots the rule in seconds and applies it to a fourth grid; many AI systems have struggled with exactly this kind of leap2.
Why it matters
It is one of the few widely watched tests where ordinary people still beat AI, so it is often quoted in arguments about how close AI is to human-like general intelligence.
Questions people ask
What does ARC-AGI stand for?
Abstraction and Reasoning Corpus for Artificial General Intelligence. The original name was the Abstraction and Reasoning Corpus; "AGI" was added later to avoid confusion with other benchmarks2.
Has any AI beaten ARC-AGI?
Not to the ARC Prize standard as of October 2026. The 2026 competition is still open, with results due on 4 December 20267. The grand prize requires open-source solutions run without internet access7.
Why is ARC-AGI hard for AI?
Each task uses a new rule the system has not seen, so memorised knowledge does not help. Chollet argues that unlimited training data lets developers "buy" skill, which hides a system's real ability to generalise1.
Related terms
Sources
- Chollet (2019), On the Measure of Intelligence, arXiv:1911.01547
- Chollet, Knoop, Kamradt and Landers (2024), ARC Prize 2024: Technical Report, arXiv:2412.04604
- ARC Prize Foundation, What is ARC-AGI?
- Chollet et al. (2025), ARC-AGI-2: A New Challenge for Frontier AI Reasoning Systems, arXiv:2505.11831
- ARC Prize Foundation, ARC-AGI-2
- ARC Prize Foundation, ARC-AGI-3
- ARC Prize Foundation, ARC Prize 2026 (key dates and rules)