Can AI agents build editable parametric CAD models in FreeCAD?

AI agents can now build editable parametric CAD models in FreeCAD often enough to be measured, but a new 100-task benchmark shows that even the best systems still miss many engineering requests once geometry, named dimensions and feature history are all checked.

The short answer: sometimes, and the best ones are improving

AI agents can build editable parametric CAD models in FreeCAD, but not reliably yet. Parametric CAD Bench v3, a benchmark released in November 2026 by Mei Chen, Yu Xia and Haipeng Mao, gave 13 complete model-plus-agent systems 100 FreeCAD tasks and scored the native files they produced. In its September 24, 2026 snapshot, the top system, Claude Code running Claude Opus 5.5, scored 61.03 percent overall. That is the best result on the board, and it still means a large share of requests were not fully met.

The benchmark's central argument is simple: an agent that writes a CAD script which runs without errors has not shown that it met the engineering request. The part also has to have the right shape, the right dimensions, and a model history an engineer can actually edit.

What makes a CAD model parametric rather than just a shape

A mesh or a plain solid is a frozen shape. A parametric model stores how the shape was built: sketches, constraints, features such as pads and pockets, and the dimensions that drive them. Change one value and the model rebuilds around it.

In FreeCAD this is done with named values and expressions. Yorik van Havre's FreeCAD manual describes giving spreadsheet cells aliases, then binding object properties and sketch constraints to those cells, so a single master value can drive the whole model. FreeCAD forbids circular dependencies, so the chain of what drives what has to be laid out cleanly. Making a variant of a part can be as simple as editing a few cells and saving under a new name.

That editability is the whole point for working engineers. A model with a named bolt circle diameter can be resized in seconds. A model that only matches the picture has to be rebuilt by hand.

How the benchmark tests agents

Version 3 has three kinds of task: 30 where the agent creates a part from a text description, 30 where it creates a part and then edits it, and 40 where it reconstructs a mechanical part from an engineering drawing. Agents work inside a pinned FreeCAD environment in a sandbox, where they can write a script, run it, read the errors and try again.

A separate program then opens the native FreeCAD file and grades it. The benchmark's developers describe two checks in their open-source validator: geometry similarity, which compares solid count, volume, surface area, bounding box and surface types against a hidden reference, and specification consistency, which checks that the key parameters exist as named, driven dimensions. In the first version they combined the two with a harmonic mean, so a perfect shape with no named parameters scores zero rather than half credit. The score is continuous rather than pass or fail, and failures count as zero.

What the results show, and what is still hard

Drawings turned out to be a relative strength for the leading system, which reached 84.66 percent on the image-to-CAD tasks. Behind it on the overall board were Claude Opus 5.5 running in the simpler mini-swe-agent harness and an OpenAI model running in Codex.

The authors stress that the numbers describe whole systems, not bare models. In the earlier version of the leaderboard, the same model scored noticeably differently depending on which agent harness drove it, and harness effects ran in both directions. The developers' explanation is that verification habits matter more than clever loops: an edit, run, print and check cycle catches CAD-specific mistakes, such as a pocket cut into empty space, that tools tuned for passing software tests tend to miss.

The hard parts are the ones a human CAD user would recognise. Getting the shape roughly right is not enough; the agent has to choose a sensible feature order, name the dimensions the request cares about, and leave a model that survives an edit. Reading orthographic drawings adds another step, turning three flat views and their dimensions into one consistent solid.

Why this matters for anyone who uses CAD

For open-source CAD, the benchmark is a useful sign: FreeCAD's scripting interface and file format are open enough to serve as a test bed, and the tasks, configurations and scoring code have been released so others can rerun them. For users, the practical lesson is to judge AI-generated parts the way the benchmark does. Open the file, check the dimensions you asked for, change one of them, and see whether the model rebuilds. A part that only looks right is a drawing, not a design.