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.

ModelsModels with a result.48
Top resultBest result on this benchmark.23.0%GPT-6 Astra
Top-3 spreadPoints from first to third.6.0 pts
YearYear of release.2026

Result and price

44 of 48 results

Results

48 results
#ModelResultThe score from the source.
01GPT-6 AstraAbove fitFar above this model's index.23.0%
12GPT-6 Sol7.0%
16Grok 4.71.0%
16Grok 4.61.0%
16Kimi K31.0%
23GPT-5.50.0%
23GLM-5.30.0%
23GPT-5.40.0%
23GLM-5.20.0%
23Grok 4.200.0%