ARC-AGI-3: people solve all of it, frontier AI scores under 1 percent
The ARC Prize Foundation's interactive benchmark is a reminder to test AI on the work you need done, not on general leaderboards.
The ARC Prize Foundation released ARC-AGI-3 in March 2026: interactive, turn-based environments in which an agent must explore, work out the goal and plan, without instructions or outside knowledge.
Human testers solved all of the environments. As of March 2026, frontier AI systems scored below 1 percent, according to the benchmark paper.
ESS view: general benchmarks say little about performance on a specific operational task. Agencies should ask vendors to score their systems on a set of the agency's own past cases.