ARC-AGI-3 debuts interactive reasoning benchmark

ARC-AGI-3 is a better proxy for agentic intelligence than another static puzzle leaderboard, because it measures learning-in-motion instead of final-answer polish. The real challenge now is not just solving the benchmark, but keeping it novel enough that models cannot train around it.

• –Relative Human Action Efficiency is the right metric here: it rewards efficient behavior, not just completion.
• –The second-best human baseline and per-level score cap make shortcut farming and speedrun exploits less valuable.
• –Replayable runs and the toolkit should accelerate RL, tool-use, and planning experiments across the ecosystem.
• –If public environments leak into training data, ARC-AGI-3 will need fresh private surfaces fast.