ARC-AGI-3 shifts benchmark into interactive games

This is the kind of benchmark shift that actually matters: it stops asking whether a model can answer cleverly once and starts asking whether an agent can operate competently over time. That makes ARC-AGI-3 less like a puzzle sheet and more like a stress test for real agent systems.

• –RHAE scores both completion and efficiency, so brute-force wandering or bloated tool loops should get punished.
• –The move from static prompts to interactive environments raises the bar on exploration policy, memory, and long-horizon planning.
• –Replayable runs and official scorecards make it useful as a development tool, not just a leaderboard.
• –Because the benchmark is built around public games and a toolkit, expect teams to compete on harness quality as much as base-model capability.
• –The biggest signal here is philosophical: ARC is betting AGI progress will show up in sustained behavior, not one-shot answer quality.