ARC-AGI-3 Launches as Frontier AI Scores Below 1%

The ARC Prize Foundation has built on the Abstraction and Reasoning Corpus introduced by François Chollet in 2019 to create the first fully interactive ARC-AGI benchmark. ARC-AGI-3 provides neither rules nor objectives, requiring agents to explore unfamiliar environments, draw inferences and plan actions on their own. It is designed to test whether models can genuinely generalize instead of relying on linguistic knowledge or memorized test questions.

The foundation formally released ARC-AGI-3 on March 25, 2026. Humans could solve every task on their first attempt, while frontier AI systems scored just 0.51% overall, with Gemini 3.1 Pro at about 0.37% and GPT-5.4 at about 0.26%. The foundation also launched ARC Prize 2026 the same day, offering more than $2 million in total prize money and setting a November 2 submission deadline.