Minebench

MineBench is a public benchmark that compares large language models by asking them to build things in a Minecraft style voxel world. Each model receives a plain language prompt, such as a wooden cabin beside a pond with a stone chimney and a small dock, and responds with raw block coordinates and block types in JSON. The site renders that output directly with no post processing and no image or 3D tooling in between. Visitors vote on anonymous head to head comparisons, choosing build A, build B, a tie, or marking both as bad. Every vote feeds a live Elo style leaderboard that ranks the models, so the rankings reflect human judgment of spatial reasoning rather than an automated score. The curated prompts are designed to probe how well a model reasons about shape, layout and structure from text alone. The site is useful to researchers, developers and enthusiasts who want a more tangible view of how frontier models differ than text benchmarks provide. A build your own feature lets signed in users generate a structure from any prompt with a chosen model, save their builds and keep their votes. The site shows a limited time offer of unlimited Gemini Flash generations with no API key needed. It does not publish a pricing page or paid plans.