ARC has released a developer preview of ARC-AGI-3, the latest version of its Artificial General Intelligence benchmark designed to test interactive reasoning. The preview makes three of six planned game-like environments publicly available, together with an API that lets researchers run their own agents against the tasks. Initial results underscore the challenge: leading AI systems scored 0% while human participants achieved 100% across the released levels. ARC said the new suite, a step up from earlier versions that focused on static reasoning, is intended to probe skills such as planning and real-time decision making that conventional deep-learning models struggle with. To spur progress, ARC opened a $10,000 contest for the first agent that can solve the current tasks. A full release of ARC-AGI-3 is scheduled for early 2026, and the group plans to incorporate community feedback gathered during the preview period.