ARC Prize Teases ARC-AGI-4: Can AI Invent on Its Own?

ARC Prize, the nonprofit AI evaluation organization, is teasing ARC-AGI-4. The next version of its benchmark will focus on autonomous open-ended innovation — testing whether AI can explore on its own, generate new ideas, and even make new inventions and discoveries.

Specific problems, scoring methods, and release timing are still under wraps. ARC Prize only says that humans still clearly lead AI at open-ended innovation, which it calls one of the most critical abilities for advancing science and technology.

The teaser also cites Dario Amodei's newly published article on slowing AI down. ARC Prize stresses that open source remains foundational to AI progress, and it opposes reducing openness or concentrating frontier AI in a handful of institutions in the name of coordinated slowdown.

ARC-AGI has long been billed as the world's hardest AGI benchmark. When ARC-AGI-3 launched, humans scored 100% while frontier AI scored just 0.51%. The latest GPT-6 Astra has closed much of that gap, but still manages only 62.7% under the unified Standard harness; it reaches 99.9% only when connected through OpenAI's own Provider Adapter.

Now ARC Prize wants to push the next challenge to whether AI can innovate on its own.

Dongcha Beating channel · Dongcha Beating community group