Announcing: ARC-AGI-3 Preview

// A North Star To AGI

Today we're announcing a public preview of ARC-AGI-3​

Intelligence is interactive.

Traditional benchmarks focus on single-turn tasks: answering a question, labeling an image, or filling in the blank. If we are going to measure human-like general intelligence, we need to go further.

Real-world intelligence unfolds over time. It depends on context, memory, and continuous learning. Humans thrive in these messy, multi-step environments. To actually evaluate AI’s general intelligence, we need interactive, multi-turn benchmarks that better capture how intelligence works.

Games are a perfect testbed. They have clear rules, clear objectives, and provide feedback after every move, but still require long-horizon planning and adaptation.

Today we are announcing our first step towards this:

• ​6 ARC-AGI-3 Preview Games - 3 are public today, 3 more arrive in August. Every game environment is novel, unique, and only requires core-knowledge priors (no language, trivia, or specialized knowledge is needed to beat the games).
  ​
• ​ARC-AGI API - Plug in any LLM, RL, or hybrid agent
  ​
• ​Agent Contest - We want your help pressure testing ARC-AGI-3. We’re hosting a 30-day $10K Agent competition in partnership with HuggingFace
  ​