ARC-AGI-3 benchmark shows humans at 100% and AI under 1%

ARC Prize just dropped ARC-AGI-3, and it makes today's frontier models look surprisingly ordinary.

While standard AI leaderboards are getting maxed out left and right, ARC Prize is calling ARC-AGI-3 the only unsaturated agentic intelligence benchmark in the world.

Humans hit 100%, AI sits below 1%

The core difference comes down to what is actually being measured.

Most current benchmarks evaluate what models already know from their pre-training data. ARC-AGI-3, by contrast, evaluates how AI systems learn new concepts on the spot.

The initial findings reveal an enormous gap. Humans score a clean 100%, while current AI models struggle to clear 1%. ARC Prize explicitly noted that this massive disparity proves artificial general intelligence has not yet been achieved, adding that the benchmark is built to stay relevant as models evolve.

Why it matters

We think ARC-AGI-3 delivers a much-needed reality check to the current AGI hype.

While frontier labs boast about models acing standardized exams, this test highlights a glaring architectural blind spot: existing systems struggle to reason when they cannot rely on memorized internet data.

For developers and researchers building autonomous agents, it provides a rare, honest target that will not saturate anytime soon. If a system claims genuine reasoning, this is the hurdle it actually has to clear.