AI Models Fail Spatial Reasoning Tests

Researchers tested AI models on 1,000 spatial puzzles, finding humans solved 98.0% while the best model, GPT-6.1 Sol, solved 83.6%. On the hardest level, people reached 94.5% accuracy versus 49.4% for the model. The study analyzed four skills, revealing that visual inputs and line tracking caused significant drops in performance across tested agents.