GPT-5.4 achieves 91% accuracy on maze-solving tasks. Gemini 3.1 Pro reaches 79%. The numbers suggest that modern language models have developed sophisticated visual-spatial reasoning — the ability to parse a maze image, identify the path, and navigate from start to finish.
Note (2026-11-14; read against the paper's full text): the essay's figures (16 model configurations, scores of 2 to 12 percent without extended reasoning, GPT-5.4 at 91 percent and Gemini 3.1 Pro at 79 percent) match an earlier statement of the results. The paper's current text evaluates 19 configurations, reports GPT-5.5 at 96 percent of the 100-maze core set, and says all configurations score only 2 to 15 percent without added reasoning budgets, with 1,096 to 22,818 tokens per solve. Its text-grid ablation, which the essay reports for Claude Sonnet 4.6 (6 to 80 percent), also shows Claude Opus 4.7 rising from 31 to 90 percent when given the correct grid.
Rodriguez Salgado (arXiv:2603.26839, March 2026) shows that the models are not solving the mazes visually. They convert the maze image to a text grid — transcribing the visual structure into a character representation — and then enumerate paths as token-level breadth-first search. The “visual planning” is actually text processing: the model transcribes what it sees into a format it can manipulate with its language processing capabilities, then applies exhaustive search in token space.
The evidence is decisive. Without extended reasoning (chain-of-thought), all model configurations drop to 2-12% accuracy. The spatial reasoning disappears entirely when the model cannot think step by step. A critical ablation isolates the bottleneck: Claude Sonnet 4.6 jumps from 6% accuracy on maze images to 80% on correctly transcribed text grids. The gap between image input and text input confirms that the difficulty is in the transcription, not the solving. Once the maze is in text form, the model solves it easily. The hard part is converting pixels to characters, not navigating the maze.
The models consume thousands of tokens per problem — far more than a spatial reasoning system would need. Each step of the breadth-first search is explicitly enumerated in the chain of thought, producing long reasoning traces that walk through the grid cell by cell. The token cost is the cost of exhaustive search in a space that a true visual planner would navigate efficiently.
The structural observation: high accuracy on a visual task does not imply visual reasoning. The models achieve accuracy through a pipeline — transcribe image to text, then search in text — that bypasses spatial comprehension entirely. The benchmark measures a composite of transcription ability and text search, and the headline number conflates the two. The 91% accuracy is real, but it is the accuracy of a lossy transcription followed by exhaustive search, not the accuracy of spatial understanding.