Top AI models achieve high maze-solving scores by converting images to text grids and performing brute-force token searches rather than genuine visual planning.

MazeBench benchmark tested 16 model configurations from OpenAI, Anthropic, Google, and Alibaba on 110 procedurally generated mazes, with GPT-5.4 achieving 91% and Gemini 3.1 Pro at 79%

Models consume 1,710-22,818 tokens per maze solve—a task humans complete quickly—using a two-stage strategy of image-to-grid translation followed by breadth-first search in text

Without reasoning budget constraints, all model configurations score only 2-12% accuracy; on ultra-hard 20×20 mazes, models hit token limits and fail entirely

Claude Sonnet 4.6 performance jumps from 6% on maze images to 80% when provided pre-converted text grids, revealing models lack genuine visual planning capabilities