Benchmark
Voxelbench
A benchmark for evaluating large language models' ability to generate voxel art and perform three-dimensional spatial reasoning, assessing how models handle spatial layout, materials and lighting in combination.
Timeline (last 90 days)
Oct 17
On Voxelbench, Claude Fable 5.5 built a complete explorable 3D world including lighting, interconnected environments and spatial relationships, using just a single prompt requesting voxel art.
Unverified50%