Learn
What is ARC-AGI-2?
Last reviewed July 16, 2026
ARC-AGI-2 is a benchmark of abstract visual puzzles: from a few example grid transformations, the solver must infer the underlying rule and apply it to a new input. Every task is novel by construction, so memorized knowledge does not help; humans solve most tasks while models score far lower. The second edition hardens the format against brute-force strategies and weighs the compute spent, so efficiency counts alongside accuracy.
Why it resists scale
Most benchmarks reward knowledge and pattern coverage, which grow with training data. ARC tasks are designed so each one requires inferring a new rule from a handful of examples: fluid intelligence rather than recall. That is why ARC scores lag far behind knowledge benchmarks even as models grow, and why progress on it is watched as a distinct signal.
Efficiency-aware scoring
ARC-AGI-2 reports not just accuracy but the cost of achieving it: a solver that brute-forces thousands of attempts per puzzle is distinguished from one that reasons its way there. That makes results harder to game with unlimited test-time compute and keeps the leaderboard anchored to something like practical intelligence per dollar.
idapt model pages carry each model's published benchmark scores, and the compare view lines several models up so generalization tests like ARC read next to the knowledge and coding lanes.
Frequently asked
See it in practice
The fastest way to understand it is to use it. Start free, no credit card.
Try idapt free