ARC Prize launched ARC-AGI-3 on March 25, calling it the world's only benchmark that current AI cannot crack. The result is striking: humans score 100%, while the best frontier models — including top versions of GPT-5.4 and Grok — score below 0.3%.

What Makes It Different

Previous ARC-AGI benchmarks were eventually saturated by AI reasoning systems, which grew powerful enough to generalize across standard public/private test splits. ARC-AGI-3 addresses this directly. The public set contains only 25 demonstration games — down sharply from prior versions — and is explicitly no longer called a "training set." Over 100 additional games make up the private evaluation set.

The benchmark places agents in interactive, game-like environments with no instructions provided. To score, a model must explore the environment, build a world model, perceive patterns, and adapt its strategy on the fly — capabilities that go well beyond pattern matching or retrieval.

Why It Matters

Co-founders François Chollet (creator of Keras) and Mike Knoop (Zapier) argue that most AI benchmarks test what models already know, not how well they learn. ARC-AGI-3 is designed to measure the latter — and current scores reveal a massive gap.

The benchmark is available now on Kaggle with over $2 million in prizes for open-source breakthroughs. ARC Prize's position is clear: until a system closes that gap, AGI remains out of reach.