# MazeBench: visual spatial reasoning in a 3D open world (MazeBench)

> Tests whether a coding agent can explore a 256-room 3D puzzle world and collect its 100 hidden gems, using box-pushing puzzles, camera rotation, and long-horizon planning.

Canonical page: https://benchlm.ai/benchmarks/mazebench

- Category: [Agentic](/agentic)
- Last updated: October 15, 2026 snapshot

## About MazeBench

- Year: 2026
- Tasks: 100 hidden gems across 256 rooms in one open world
- Format: Gems collected out of 100, with rooms visited as a second score
- Difficulty: Long-horizon spatial planning, with single puzzles that can take over one hundred moves
- Paper: [MazeBench Results](https://mazebench.com/blog?post=maze-bench-results)

MazeBench is a three-dimensional world of 256 rooms built around Sokoban-style box puzzles, with elevator tiles, toggleable walls, and icy slopes. The agent sees the world through a camera it can rotate to 20 angles, and 100 gems are hidden across the rooms. Agents play inside their own coding-agent harnesses, such as Codex and Claude Code, through an MCP server with eleven actions: four movement, four camera, undo, level reset, and teleport to a room already visited. Runs have no action limit and are stopped when the agent keeps circling without finding new board states. The leaderboard covers ASCII mode, where each tile is a 4x4 block of characters, with Python tools either enabled or disabled. The engine, level editor, and solution algorithms are open source.

MazeBench is currently displayed on BenchLM for reference, but it is excluded from the weighted scoring formula.

## Leaderboard (62 published runs)

| Rank | Model | Configuration | Creator | Score |
|------|-------|---------------|---------|-------|
| 1 | [GPT-6 Astra](/models/gpt-6-astra) | Codex · max reasoning · Tools · 113 rooms · Paused snapshot | OpenAI | 23% |
| 2 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | Codex · max reasoning · Tools · 102 rooms · Paused snapshot | OpenAI | 17% |
| 3 | [GPT-6 Astra](/models/gpt-6-astra) | Codex · max reasoning · No tools · 104 rooms · Paused snapshot | OpenAI | 16% |
| 4 | [Claude Opus 5.5](/models/claude-opus-5-5) | Claude Code · max reasoning · Tools · 68 rooms · Stopped snapshot | Anthropic | 13% |
| 5 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | Codex · max reasoning · Tools · 67 rooms | OpenAI | 13% |
| 6 | [Gemini 4 Argon](/models/gemini-4-argon) | Antigravity · high reasoning · No tools · 82 rooms · In-progress snapshot | Google | 12% |
| 7 | [Claude Opus 5](/models/claude-opus-5) | Claude Code · max reasoning · Tools · 62 rooms | Anthropic | 12% |
| 8 | [Claude Fable 5](/models/claude-fable) | Claude Code · max reasoning · Tools · 43 rooms | Anthropic | 11% |
| 9 | [Claude Fable 5.1](/models/claude-fable-5-1) | Claude Code · max reasoning · Tools · 33 rooms | Anthropic | 10% |
| 10 | [GPT-6.1 Sol](/models/gpt-6-1-sol) | Codex · max reasoning · No tools · 66 rooms · Paused snapshot | OpenAI | 9% |
| 11 | [Claude Opus 5.5](/models/claude-opus-5-5) | Claude Code · max reasoning · No tools · 50 rooms · In-progress snapshot | Anthropic | 9% |
| 12 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | Claude Code · max reasoning · Tools · 49 rooms · Stopped snapshot | Anthropic | 8% |
| 13 | [GPT-6 Sol](/models/gpt-6-sol) | Codex · max reasoning · Tools · 38 rooms · Paused snapshot | OpenAI | 7% |
| 14 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | Codex · max reasoning · Tools · 12 rooms | OpenAI | 6% |
| 15 | [Gemini 3.8 Flash](/models/gemini-3-8-flash) | Google Antigravity · high reasoning · Tools · 12 rooms | Google | 4% |
| 16 | [Claude Fable 5.1](/models/claude-fable-5-1) | Claude Code · max reasoning · No tools · 12 rooms · Paused snapshot | Anthropic | 2% |
| 17 | [Claude Fable 5](/models/claude-fable) | Claude Code · max reasoning · No tools · 10 rooms | Anthropic | 1% |
| 18 | [Muse Wonder Full](/models/muse-wonder-full) | Muse Code · max reasoning · No tools · 8 rooms · In-progress snapshot | Meta | 1% |
| 19 | [Claude Sonnet 5.5](/models/claude-sonnet-5-5) | Claude Code · max reasoning · No tools · 8 rooms · Stopped snapshot | Anthropic | 1% |
| 20 | [Claude Opus 5](/models/claude-opus-5) | Claude Code · max reasoning · No tools · 8 rooms | Anthropic | 1% |
| 21 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | Claude Code · max reasoning · Tools · 7 rooms · In-progress snapshot | Anthropic | 1% |
| 22 | [GPT-6 Sol](/models/gpt-6-sol) | Codex · max reasoning · No tools · 7 rooms · Paused snapshot | OpenAI | 1% |
| 23 | [Grok 4.7](/models/grok-4-7) | Grok Build · xhigh reasoning · No tools · 7 rooms · Stopped snapshot | xAI | 1% |
| 24 | [Grok 4.6](/models/grok-4-6) | Grok Build · xhigh reasoning · Tools · 6 rooms | xAI | 1% |
| 25 | [Claude Haiku 5.5](/models/claude-haiku-5-5) | Claude Code · max reasoning · No tools · 5 rooms · Stopped snapshot | Anthropic | 1% |
| 26 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | Antigravity · high reasoning · Tools · 5 rooms | Google | 1% |
| 27 | [Gemini 3.7 Flash](/models/gemini-3-7-flash) | Antigravity · high reasoning · Tools · 4 rooms | Google | 1% |
| 28 | [GPT-5.6 Sol](/models/gpt-5-6-sol) | Codex · max reasoning · No tools · 4 rooms | OpenAI | 1% |
| 29 | [Kimi K3](/models/kimi-k3) | Kimi Code · max reasoning · Tools · 4 rooms | Moonshot AI | 1% |
| 30 | [GPT-5.5](/models/gpt-5-5) | Codex · xhigh reasoning · No tools · 4 rooms | OpenAI | 0% |
| 31 | [Claude Opus 4.8](/models/claude-opus-4-8) | Claude Code · max reasoning · No tools · 4 rooms | Anthropic | 0% |
| 32 | [Muse Wonder](/models/muse-wonder) | Muse Code · max reasoning · No tools · 4 rooms · Stopped snapshot | Meta | 0% |
| 33 | [Muse Spark 1.3](/models/muse-spark-1-3) | Muse Code · max reasoning · No tools · 3 rooms · Paused snapshot | Meta | 0% |
| 34 | [GPT-6 Luna](/models/gpt-6-luna) | Codex · max reasoning · No tools · 2 rooms · Stopped snapshot | OpenAI | 0% |
| 35 | [Kimi K3](/models/kimi-k3) | Kimi Code · high reasoning · No tools · 2 rooms | Moonshot AI | 0% |
| 36 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | Codex · max reasoning · Tools · 2 rooms | OpenAI | 0% |
| 37 | [GPT-5.6 Terra](/models/gpt-5-6-terra) | Codex · max reasoning · No tools · 2 rooms | OpenAI | 0% |
| 38 | [GLM-5.3-Flash](/models/glm-5-3-flash) | Claude Code · max reasoning · Tools · 2 rooms | Z.AI | 0% |
| 39 | [Claude Sonnet 4.6](/models/claude-sonnet-4-6) | Claude Code · high reasoning · No tools · 2 rooms | Anthropic | 0% |
| 40 | [Gemini 3.6 Flash](/models/gemini-3-6-flash) | Antigravity · high reasoning · Tools · 2 rooms | Google | 0% |
| 41 | [deepseek-v4-pro · Prime Agent high · Tools](https://mazebench.com/leaderboard/ai?mode=ascii&tools=1&metric=gems&run=2026-08-25T10-46-04-610-d58bda) | Prime Agent · high reasoning · Tools · 2 rooms | DeepSeek | 0% |
| 42 | [GLM-5.3](/models/glm-5-3) | Prime Agent · high reasoning · Tools · 2 rooms | Z.AI | 0% |
| 43 | [GPT-5.4](/models/gpt-5-4) | Codex · xhigh reasoning · No tools · 2 rooms | OpenAI | 0% |
| 44 | [Qwen3.8 Max](/models/qwen3-8-max) | Prime Agent · high reasoning · Tools · 2 rooms | Alibaba | 0% |
| 45 | [Claude Sonnet 5](/models/claude-sonnet-5) | Claude Code · max reasoning · No tools · 1 rooms | Anthropic | 0% |
| 46 | [Muse Spark 1.2](/models/muse-spark-1-2) | Muse Code · xhigh reasoning · No tools · 1 rooms · Paused snapshot | Meta | 0% |
| 47 | [Claude Sonnet 5](/models/claude-sonnet-5) | Claude Code · max reasoning · Tools · 1 rooms | Anthropic | 0% |
| 48 | [Muse Spark 1.3](/models/muse-spark-1-3) | Muse Code · ultra reasoning · Tools · 1 rooms | Meta | 0% |
| 49 | [Gemini 3.1 Pro](/models/gemini-3-1-pro) | Codex · high reasoning · No tools · 1 rooms | Google | 0% |
| 50 | [GPT-5.6 Luna](/models/gpt-5-6-luna) | Codex · max reasoning · No tools · 1 rooms | OpenAI | 0% |
| 51 | [GLM-5.2](/models/glm-5-2) | Prime · high reasoning · No tools · 1 rooms | Z.AI | 0% |
| 52 | [MiniMax M3](/models/minimax-m3) | Prime · high reasoning · No tools · 1 rooms | MiniMax | 0% |
| 53 | [Gemini 3.5 Flash](/models/gemini-3-5-flash) | Prime · high reasoning · No tools · 1 rooms | Google | 0% |
| 54 | [Claude Haiku 4.5](/models/claude-haiku-4-5) | Claude Code · No tools · 1 rooms | Anthropic | 0% |
| 55 | [deepseek-v4-pro · Prime high · No tools](https://mazebench.com/leaderboard/ai?mode=ascii&tools=0&metric=gems&run=2026-07-19T04-09-40-388-f36c15) | Prime · high reasoning · No tools · 1 rooms | DeepSeek | 0% |
| 56 | [GPT-5.3-Codex-Spark](/models/gpt-5-3-codex-spark) | Codex · xhigh reasoning · No tools · 1 rooms | OpenAI | 0% |
| 57 | [deepseek-v4-flash · Prime high · No tools](https://mazebench.com/leaderboard/ai?mode=ascii&tools=0&metric=gems&run=2026-07-19T02-30-42-219-3cd147) | Prime · high reasoning · No tools · 1 rooms | DeepSeek | 0% |
| 58 | [Grok 4.20](/models/grok-4-20-beta) | Prime · high reasoning · No tools · 1 rooms | xAI | 0% |
| 59 | [Nemotron 3 Ultra](/models/nemotron-3-ultra) | Prime · high reasoning · No tools · 1 rooms | NVIDIA | 0% |
| 60 | [DeepSeek V4 Flash 0731](/models/deepseek-v4-flash-0731) | Prime Agent · high reasoning · Tools · 1 rooms | DeepSeek | 0% |
| 61 | [MiMo-V2.5-Pro](/models/mimo-v2-5-pro) | Prime · high reasoning · No tools · 1 rooms | Xiaomi | 0% |
| 62 | [Llama 4 Maverick](/models/llama-4-maverick) | Prime · high reasoning · No tools · 1 rooms | Meta | 0% |

## FAQ

### What does MazeBench measure?

MazeBench measures whether an agent can plan through a three-dimensional puzzle world. The agent explores 256 rooms of box-pushing puzzles through a camera it can rotate, and scores one percent for each of the 100 hidden gems it collects. Rooms visited is a second score. The published runs use the text-based ASCII mode.

### Why do some rows say Tools and others No tools?

Tools means the agent could run Python, which lets it write a solver for the puzzles. No tools means it had to plan inside the game. The same model can have one run in each setting, and the two scores are not interchangeable. A model's page shows its best published run and names the setting.

### Why is MazeBench display only?

Each result is one game played through a specific harness, reasoning setting, and tool setting, and many are snapshots of runs that were paused, stopped at a usage limit, or still going. The scores are useful for comparing those exact runs, but they are not a model-only measurement, so they stay out of BenchLM's rankings.