ARC-AGI-3 exposes open source lag

ARC-AGI is less a benchmark than a trapdoor: it rewards models that can adapt, not models that merely sound competent. That is why the open-vs-closed gap looks much bigger here than on friendlier public leaderboards. ARC-AGI-2 explicitly says scale alone is not enough, brute-force search is not intelligence, and its calibrated private evals are meant to resist benchmark-maxing. ARC-AGI-3 shifts from static puzzles to interactive environments with exploration, planning, memory, goal acquisition, and continuous adaptation. The gap is partly about scaffolding: closed labs can bundle reasoning loops, tool use, and test-time refinement around frontier models, while many open releases ship the base weights alone. ARC Prize's latest analysis frames the remaining gap as engineering for raw score gains and ideas for efficiency gains, which is why ARC is a better monitor of the "real" frontier than static public leaderboards.