ARC-AGI-3 Launches: Humans Score 100%, Frontier AI Models Score Below 1%
A new benchmark designed to measure genuine novel reasoning — not pattern matching — finds that the best AI systems fail almost completely on tasks any human can solve. The gap is the point.
The ARC Prize Foundation released ARC-AGI-3 on Wednesday, a benchmark that tests AI systems on their ability to learn and reason in interactive environments they've never encountered before. The results are stark: humans score 100%, while frontier AI models score below 1%. As @arcprize announced, it's positioned as "the only unsaturated agentic intelligence benchmark" — meaning no AI system has come close to saturating it, unlike older benchmarks that models have effectively solved through scale and training data coverage.
Unlock the full briefing
Get every story in today's briefing, the full archive, and the daily AI intelligence brief.
All stories today
Full archive
Daily brief
Cancel anytime. Payments powered by Stripe.