Skip to content

MazeBench: visual spatial reasoning in a 3D open world

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.

42 models

Higher is better.

  1. 1GPT-6 AstraOpenAI23
  2. 2GPT-6.1 SolOpenAI17
  3. 3Claude Opus 5.5Anthropic13
  4. 4GPT-5.6 SolOpenAI13
  5. 5Gemini 4 ArgonGoogle12
  6. 6Claude Opus 5Anthropic12
  7. 7Claude Fable 5Anthropic11
  8. 8Claude Fable 5.1Anthropic10
  9. 9Claude Sonnet 5.5Anthropic8
  10. 10GPT-6 SolOpenAI7
  11. 11GPT-5.6 TerraOpenAI6
  12. 12Gemini 3.8 FlashGoogle4
  13. 13Muse Wonder FullMeta1
  14. 14Kimi K3Moonshot AI1
  15. 15Claude Haiku 5.5Anthropic1
  16. 16Grok 4.6xAI1
  17. 17Gemini 3.7 FlashGoogle1
  18. 18Gemini 3.1 ProGoogle1
  19. 19Grok 4.7xAI1
  20. 20Muse WonderMeta0
  21. 21Claude Opus 4.8Anthropic0
  22. 22Qwen3.8 MaxAlibaba0
  23. 23GPT-5.5OpenAI0
  24. 24GLM-5.3Z.AI0
  25. 25GPT-5.4OpenAI0
  26. 26GPT-6 LunaOpenAI0
  27. 27Muse Spark 1.2Meta0
  28. 28Claude Sonnet 5Anthropic0
  29. 29GPT-5.6 LunaOpenAI0
  30. 30Gemini 3.6 FlashGoogle0
  31. 31Gemini 3.5 FlashGoogle0
  32. 32GLM-5.2Z.AI0
  33. 33GLM-5.3-FlashZ.AI0
  34. 34Grok 4.20xAI0
  35. 35Claude Sonnet 4.6Anthropic0
  36. 36MiniMax M3MiniMax0
  37. 37MiMo-V2.5-ProXiaomi0
  38. 38Nemotron 3 UltraNVIDIA0
  39. 39Claude Haiku 4.5Anthropic0
  40. 40Llama 4 MaverickMeta0
Higher is better. Showing the top 40 of 42. Bar colour: blue proprietary, orange open weights, green other.

About this benchmark

Metric
Gems collected out of 100, with rooms visited as a second score
Tasks
100 hidden gems across 256 rooms in one open world
Authors
Jonathan Pappas and David Pappas (2026)

Scores are as reported by model developers or evaluators and may use different settings (for example reasoning effort or tools). Compare with care.

All results

1GPT-6 Astra OpenAI23
2GPT-6.1 Sol OpenAI17
3Claude Opus 5.5 Anthropic13
4GPT-5.6 Sol OpenAI13
5Gemini 4 Argon Google12
6Claude Opus 5 Anthropic12
7Claude Fable 5 Anthropic11
8Claude Fable 5.1 Anthropic10
9Claude Sonnet 5.5 Anthropic8
10GPT-6 Sol OpenAI7
11GPT-5.6 Terra OpenAI6
12Gemini 3.8 Flash Google4
13Muse Wonder Full Meta1
14Kimi K3 Moonshot AI1
15Claude Haiku 5.5 Anthropic1
16Grok 4.6 xAI1
17Gemini 3.7 Flash Google1
18Gemini 3.1 Pro Google1
19Grok 4.7 xAI1
20Muse Wonder Meta0
21Claude Opus 4.8 Anthropic0
22Qwen3.8 Max Alibaba0
23GPT-5.5 OpenAI0
24GLM-5.3 Z.AI0
25GPT-5.4 OpenAI0
26GPT-6 Luna OpenAI0
27Muse Spark 1.2 Meta0
28Claude Sonnet 5 Anthropic0
29GPT-5.6 Luna OpenAI0
30Gemini 3.6 Flash Google0
31Gemini 3.5 Flash Google0
32GLM-5.2 Z.AI0
33GLM-5.3-Flash Z.AI0
34Grok 4.20 xAI0
35Claude Sonnet 4.6 Anthropic0
36MiniMax M3 MiniMax0
37MiMo-V2.5-Pro Xiaomi0
38Nemotron 3 Ultra NVIDIA0
39Claude Haiku 4.5 Anthropic0
40Llama 4 Maverick Meta0
41DeepSeek V4 Flash 0731 DeepSeek0
42Muse Spark 1.3 Meta0