Launch: ARC-AGI-3 Toolkit

// ARC-AGI-3 Toolkit

Today we’re releasing the ARC-AGI-3 Toolkit.

Your agents can now interact with ARC-AGI-3 environments at >2,000 FPS locally.

We're open-sourcing the environment engine, 3 human-verified games (AI scores <5%), and human baseline scores.

ARC-AGI-3 moves from static puzzles to interactive environments.

Agents now take sequential actions inside environments, letting us measure how efficiently they learn, not just whether they succeed.

The toolkit open sources the ARC-AGI-3 game engine which runs locally at >2,000 FPS, giving developers a >250x speedup over our API.

Get started in 5 lines of code.

We’re also releasing:

• 3 human-verified games - The first official Public Set games. Human-verified solvable with improved onboarding. Current AI scores sit below 5%.
  ​
• Human-normalized scoring - We’re adopting Relative Human Action Efficiency as our new scoring methodology. All official scores will be normalized to human baselines. Read more

​
ARC-AGI-3 launches March 25, 2026.