Skip to content

Abstraction and Reasoning Corpus for AGI v3

An interactive successor to ARC-AGI-2 that evaluates whether an AI agent can learn unfamiliar task mechanics through action and feedback.

22 models

Higher is better.

  1. 1GPT-6 AstraOpenAI62.7
  2. 2GPT-6.1 SolOpenAI52.7
  3. 3Gemini 4 ArgonGoogle46.3
  4. 4Claude Opus 5Anthropic30.2
  5. 5Gemini 3.8 FlashGoogle10.4
  6. 6GPT-5.6 SolOpenAI7.8
  7. 7GPT-6 SolOpenAI4.6
  8. 8Grok 4.6xAI2.1
  9. 9Grok 4.7xAI1.8
  10. 10Claude Opus 4.8Anthropic1.5
  11. 11GPT-5.6 TerraOpenAI0.8
  12. 12Claude Opus 4.6Anthropic0.5
  13. 13Claude Haiku 5.5Anthropic0.5
  14. 14GPT-5.5OpenAI0.4
  15. 15Gemini 3.1 ProGoogle0.4
  16. 16Grok 4.5xAI0.3
  17. 17GPT-5.4OpenAI0.2
  18. 18Claude Opus 4.7 (Adaptive)Anthropic0.2
  19. 19Claude Opus 4.7Anthropic0.2
  20. 20GPT-5.6 LunaOpenAI0.2
  21. 21GPT-6 LunaOpenAI0.1
  22. 22Grok 4.20xAI0.1
Higher is better. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Agentic task completion under a capped evaluation budget
Tasks
Interactive game-like tasks with hidden rules
Authors
ARC Prize Foundation (2026)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.