Today we're releasing a developer preview of our next-gen benchmark, ARC-AGI-3.

The goal of this preview, leading up to the full version launch in early 2026, is to collaborate with the community. We invite you to provide feedback to help us build the most robust and effective benchmark possible.

ARC-AGI-3 is the continuation of the ARC series of AI benchmarks:

• Focus on generalization and adaptation to novelty
• Easy for humans, yet extremely difficult for AI
• Built exclusively on Core Knowledge priors, with no other domain-specific knowledge required

The biggest evolution is a complete shift to an interactive format. ARC 3 is a collection of unique, novel games set in the ARC grid world. To succeed, an AI must learn on the fly through efficient trial and error. There are no instructions, so everything is up for you to figure out:

• What are the underlying concepts and mechanics?
• What do the controls do?
• What is the goal, and how do I achieve it?

We are directly probing an AI's ability to efficiently explore, learn, and plan when faced with a completely unknown task. So far, all systems we've tested score 0. Yet all these games are fully solvable by humans in a few minutes with no prior training.

点击图片查看原图