The ARC Prize Foundation will release ARC-AGI-3 on March 25, marking the first major format change to the benchmark since François Chollet introduced ARC in 2019. Gone are the static grid puzzles. Version 3 drops agents into video-game-like environments where they have to explore, plan, and adapt with zero instructions.
The benchmark includes over 1,000 levels across 150-plus hand-crafted environments, each designed so memorization won't help. Every environment is human-solvable, but the scoring metric isn't pass/fail. It measures action efficiency: how many steps does a system need to reach a goal compared to a human? ARC Prize Foundation president Greg Kamradt has called this a way to formally compare human versus AI learning efficiency for the first time. Three preview environments are already playable, and a developer toolkit shipped in late January, letting researchers run environments locally at up to 2,000 FPS.
The original source material claims Y Combinator will host a launch event featuring Kamradt, Chollet, and OpenAI CEO Sam Altman, though ARC Prize's own site lists only the public benchmark release for March 25. Altman has previously collaborated with the ARC Prize team, joining the o3 announcement livestream in late 2024, and has expressed interest in partnering on future benchmarks.
ARC-AGI-3 arrives at an odd moment. AI reasoning models have gotten dramatically better at static benchmarks, with the top ARC Prize 2025 competition score hitting 24% on ARC-AGI-2's private dataset at just $0.20 per task. But the foundation's position hasn't changed: current AI reasoning is still tightly coupled to model knowledge in ways human reasoning is not. The interactive format is built to expose exactly that gap.
Bottom Line
ARC-AGI-3 shifts from static puzzles to 1,000+ interactive game environments that measure how efficiently AI learns compared to humans, launching March 25.
Quick Facts
- Launch date: March 25, 2026
- 1,000+ levels across 150+ hand-crafted environments
- First major format change since ARC was introduced in 2019
- Developer toolkit released January 29, 2026
- Top ARC Prize 2025 score: 24% on ARC-AGI-2 private dataset at $0.20/task
- Sam Altman attendance at launch event: unverified
