I built an eval to measure the ability of various LLMs to create CAD. Gemini 2.5 and o3 are leading the pack.
It looks like reasoning may be a big enabler. Anecdotally, I've noticed that the reasoning models effectively think through how they're going to make the part.