ARC Prize: ARC-AGI-3 Competition
Build an agent that works out and beats unfamiliar puzzle games with no rules
In the ARC-AGI-3 competition you build an AI agent that plays small turn-based puzzle games it has never seen. The agent gets no instructions, rules or stated goal. It has to explore, work out what winning means and then win. Hidden games score the submissions, and humans can solve all of them.
Free · Login needed · API · Advanced
About ARC Prize: ARC-AGI-3 Competition
It trains agent design, exploration strategies, program synthesis and the careful use of models under compute limits. It suits advanced programmers and machine learning researchers.
Submissions run as Kaggle notebooks with no internet access, and an SDK lets you test against public games. Entry is free with a Kaggle account. Prize money requires open-sourcing your solution. The 2026 round closes for submissions on 2 November 2026.
Made this game? Claim it →
Tips and discussion
No tips yet. Solved it? Share how, and wrap answers in ||spoilers||.