P3D-Bench - Assembly-3D Judge: leaderboard
P3D-Bench (2026) Assembly-3D on 203 annotated multi-part assemblies: an MLLM judge (Gemini 3.1 Pro) rates the rendered CadQuery or OpenSCAD assembly for semantic, geometric and aesthetic similarity. Programs are compiled, meshed and aligned to the ground truth, invalid outputs take the worst value, and each general model runs with its maximum thinking budget; the score is on a 0-100 scale averaged over output formats, higher is better.
Metric: MLLM-judge bucket score (0-100) (Gemini 3.1 Pro rates rendered views for semantic, geometric and aesthetic similarity and answers specification-derived QA) on 203 annotated multi-part assemblies (CadQuery and OpenSCAD outputs), the model writes a parametric CAD program that is compiled, meshed and aligned to the ground truth; invalid outputs take the worst value; each general model runs with its maximum thinking budget; averaged over the output formats; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.5 (xHigh) | 54.1 | #49 (GPT-5.5) |
| 2 | Gemini 3.1 Pro (Preview) (High) | 50.7 | #94 (Gemini 3.1 Pro (Preview)) |
| 3 | Claude Opus 4.6 (Max) | 37.6 | #102 (Claude Opus 4.6) |
| 4 | Qwen 3.6 Plus (Thinking) | 16.3 | #222 (Qwen 3.6 Plus) |
| 5 | GLM-5V Turbo | 12.9 | #344 |
| 6 | Seed 2.0 Pro (High) | 12.9 | #211 (Seed 2.0 Pro) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=p3d-bench-assembly-3d-judge · How It Works · Data refreshed daily, snapshot 2026-11-13.