ARC-AGI-3 is here. The ARC Prize team built a new benchmark around 135 interactive visual games designed to measure learning efficiency, not pattern recall. The core problem it targets: existing benchmarks are saturated, meaning models have effectively maxed them out without achieving anything resembling general intelligence.
The distinction the piece makes is worth your attention. Current tests reward memorization. ARC-AGI-3 rewards adaptation, specifically the gap between how fast humans learn new tasks versus how current models perform under the same conditions. It also introduces tool-enabled, real-world evaluation formats, a structural shift away from static question-answer formats.
The deeper argument is about research direction. If benchmarks are saturated, labs optimize for the benchmark, not the capability. ARC-AGI-3 is an attempt to reset that incentive. Read the full coverage for the breakdown of why benchmark design is now one of the most consequential decisions in AI development.
[WATCH ON YOUTUBE →]