Maze-Bench: leaderboard
Evaluates whether a model can read a rectangular maze image and produce an escape route, measuring continuous visual and spatial reasoning. Each task is a 2048x2048 maze image with no coordinates, grid labels or textual representation; the model returns a single line of movement commands relative to its current heading — forward, backward, right and left — generated in one response with no intermediate feedback, and any extra text is a format error. Each maze scores 100 x P x E, where progress P is m/(m+r) and efficiency E is D/(m+r), with D the shortest solution length, m the moves completed without hitting a wall and r the remaining distance from the final position. The suite is five mazes at each of six sizes from 4x4 to 18x18; the mean becomes official only once all 30 are graded, and partial coverage is shown as a provisional mean instead.
Metric: Mean Score. Source: huggingface.co. Saturation forecast: Around November 2027. 44 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-6 Astra (Medium) | 94.99 | #6 (GPT-6 Astra) |
| 2 | GPT-6.1 Sol (Medium) | 88.57 | #7 (GPT-6.1 Sol) |
| 3 | Claude Opus 5.5 (Medium) | 85.17 | #3 (Claude Opus 5.5) |
| 4 | GPT-6.1 Terra (Medium) | 79.63 | #15 (GPT-6.1 Terra) |
| 5 | Kimi K3.1 Pro (High) | 73.09 | #17 (Kimi K3.1 Pro) |
| 6 | Claude Sonnet 5.5 (Medium) | 58.47 | #14 (Claude Sonnet 5.5) |
| 7 | Claude Fable 5.1 (Medium) | 55.35 | #10 (Claude Fable 5.1) |
| 8 | GPT-6 Sol (Medium) | 54.65 | #19 (GPT-6 Sol) |
| 9 | Claude Haiku 5.5 (Medium) | 52.65 | #43 (Claude Haiku 5.5) |
| 10 | Claude Opus 5 (Medium) | 50.86 | #16 (Claude Opus 5) |
| 11 | Kimi K3 (High) | 48.14 | #35 (Kimi K3) |
| 12 | Gemini 3.8 Flash (Medium) | 40.67 | #30 (Gemini 3.8 Flash) |
| 13 | Claude Fable 5 (Medium) | 39.35 | #20 (Claude Fable 5) |
| 14 | Kimi K3 (Max) | 32.89 | #35 (Kimi K3) |
| 15 | GPT-5.6 Sol (Medium) | 31.46 | #22 (GPT-5.6 Sol) |
No result here: #9 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=maze-bench · How It Works · Data refreshed daily, snapshot 2026-11-05.