ARC-AGI Leaderboard

The ARC-AGI-3 leaderboard tracks AI systems' ability to solve novel interactive tasks while measuring efficiency through cost-per-task performance. The benchmark has evolved from testing passive intelligence to requiring AI agents to adapt dynamically to new environments, with solutions categorized as reasoning systems, base language models, or competition-grade systems operating under computational constraints. Key insights show how increased reasoning time affects performance and demonstrate that true intelligence requires solving problems efficiently with minimal resources.