Frontier models fail ARC-AGI-3 reasoning test

ARC-AGI-3 is a reality check for the scaling hypothesis—true intelligence requires efficient adaptation to novel environments, not just vast knowledge retrieval.

• –Interactive levels replace static grids to eliminate benchmark contamination and force real-time exploration
• –Humans solve 100% of tasks effortlessly, while frontier models like Gemini 3 and GPT-5.4 show near-zero action efficiency
• –Reasoning trace analysis reveals that previous high scores on ARC were likely due to the presence of ARC-like data in model training sets
• –A new efficiency-based scoring protocol penalizes agents that cannot convert feedback into strategies as quickly as humans
• –A high-performance local training toolkit (2,000 FPS) has been released to help researchers bridge the reasoning gap