ARC-AGI-3

Games with no instructions. Each is a small world on a 64×64 screen in sixteen colours; you are not told the rules, only which actions you have (four directions, interact, click a point, undo, reset). Clear its levels in as few actions as a person would. This is the ARC Prize 2026 benchmark, on its own official toolkit and games, played offline.

Scoring

ARC's own RHAE: for each level you clear, (the human baseline's actions ÷ your actions)², capped at 1.15, weighted by the level's number; a game's score is that as a percentage of all its levels. The check plays the season's five games, up to 400 actions each, and your score is the mean. Uncleared levels score 0; most agents score close to that.

Your agent

Exactly an ARC-AGI-3-Agents agent: agent.py with a class Agent that has choose_action(frames, latest_frame) (and, if you like, is_done(frames, latest_frame)). latest_frame is the toolkit's FrameData: the frame layers, state, levels_completed, win_levels and available_actions. Return a GameAction, or (GameAction.ACTION6, {"x": .., "y": ..}) for a click. Your agent plays through the official toolkit, in its own process, against the referee's local server; it never sees a game's source.

The games, season 1

LS20, FT09, VC33, SP80 and LP85, from the public set (referee/games.txt, files in referee/environment_files/).

Credits

The toolkit, the engine and the games are © 2026 ARC Prize Foundation, under the MIT licence, kept with their notices. ARC Prize and ARC-AGI are the Foundation's names; this project is not affiliated with it. The referee and the viewer, which draws frames in the official palette, are this project's, under MIT.