Spatial & 3D building

Vision2Code: four orthographic views in, a CadQuery program out. The metric is voxel IoU — the executed program's solid compared voxel-by-voxel against the ground-truth part; programs that fail to execute score 0 (IoU-score = voxel IoU × exec rate). In the with tools setting the model gets a Python sandbox to render, measure and iterate before submitting. The two settings are different evaluations and are never merged.

With tools agentic Python sandbox

* BenchCAD lists Claude Fable 5.1 at 92.6% with tools (Anthropic's Opus 5.5 system card, fig. 8.13.2.A). OpenAI's GPT-6 Astra launch post quotes 84.3% for the same model — we show the benchmark's number and keep both contexts visible. Lab-reported rows are vendor self-reported voxel IoU, not re-graded by BenchCAD; Anthropic's figures use a random 1,000-file subset, so their split is not the same as the no-tools column beside them.

No tools single shot, IoU-score × execution rate

GPT-6 Astra has no published no-tools score, so it appears only in the with-tools table. Data: BenchCAD (CC BY 4.0), snapshot Oct 9, 2026. Machine-readable source.

Models build Minecraft-style voxel structures from a text prompt. There is no objectively correct answer: visitors compare two anonymous builds and vote, and a global Bradley–Terry model turns those votes into a rating on a 400-point Elo scale with a 95% confidence interval. Snapshot of the public leaderboard, Oct 9, 2026.

Top 15 of 72 ranked models in our snapshot. The live board moves with every vote.

Real builds the top five models' three strongest prompts — drag to orbit, pinch or scroll to zoom

The builds are the byte-exact artifacts MineBench serves its own visitors (checks SHA-verified), shown with flat colors — MineBench's Faithful texture pack is separately licensed and not redistributed. Block colors are our approximation. Builds carry no stated license on the host; the MineBench repo is MIT. Credit: MineBench by Ammaar Alam.

• BenchCAD — programmatic CAD benchmark (Vision2Code, Vision QA, Code QA, Code Edit). Scores from the public leaderboard JSON; benchmark data is CC BY 4.0, code MIT. Rows marked “Lab-reported” are vendor self-reports, not re-graded; others are BenchCAD's own runs. Leaderboard · Repo · Dataset · snapshot Oct 9, 2026
• MineBench by Ammaar Alam — voxel building arena, blind human votes scored with a global Bradley–Terry model. Leaderboard order and the 15 builds shown are a one-time manual snapshot from the public pages, credited per build; hosted outputs carry no stated license, the repo is MIT. Site · Leaderboard · Repo (MIT) · snapshot Oct 9, 2026

Both snapshots refresh manually, on occasion — they are not part of the daily pipeline. Model names link to our catalog pages only where the benchmark's name matches a cataloged model.