It's the first time ever I am submitting a link to my AI 3D Modeling software (even if it's just a blog post), so I would appreciate any feedback - or request for other data you may want to see.
I'm not being nit picky here. I think there are issues beyond terminology that you may not be familiar with, as it is clearly not your field. That's ok.
The "design" in computer aided design is engineering design. This is not the same definition of "design" used in, say, graphic design. Something is not called CAD because it helps you create an image that looks like a product on a computer. It is CAD because it creates engineering design files (blueprints) that can be used for the physical manufacture of a device. This places very tight and important constraints on the methods used, and capabilities supported.
Blender is a sculpting program. Its job is to create geometry that can be fed into a rendering program to make pretty pictures. Parasolid is a CAD geometry kernel at the core of many CAD programs, which has the job of producing manufacturable blueprints. The operations supported map to physical manufacturing steps - milling, lathe, and drill operations. The modeling steps use constraints in order to make sure, e.g., that screw holes line up. Blender doesn't support any of that.
To an engineer, saying that an LLM gave you a blender script for a CAD operation is causing all sorts of alarm klaxons to go off.
If Blender can export a .3mf file format and slicer gets it ready for 3D printing (gcode that actually instructs the printer head). Is the slicer actually CAD software?
And if you can export many formats that work with some manufacturing devices and you built a model in blender, did blender not help you with CAD?
Results are amazing! 2.5 and 3 seems way way head.
2.5 stands between GPT-5 and GPT-5.1, where GPT-5 is the best of the 3.
In preliminary evals Gemini 3 seems to be way better than all, but I will know when I run extended benchmarks tonight.
I guess openscad would be a sweet spot in the middle. Good shout, might experiment.
It generated a blender script that makes the model.
I will run 80 3D model generations benchmark tomorrow and update this comment with the results about cost/speed/quality.