New ARC-AGI-3 benchmark reveals major gap between human and AI reasoning, with frontier systems scoring below 1% compared to 100% human success as of March 2026.

ARC-AGI-3 is an interactive benchmark designed to test agentic intelligence through abstract, turn-based environments requiring exploration, goal inference, and planning without explicit instructions

The benchmark avoids language and external knowledge, focusing solely on fluid adaptive efficiency on novel tasks using only Core Knowledge priors

Humans achieve 100% success rate on ARC-AGI-3 environments while frontier AI systems score below 1% as of March 2026, highlighting a significant capability gap

Difficulty levels are calibrated through extensive human testing, with scoring based on an efficiency framework grounded in human action baselines