I asked Claude and GPT-5.5 to design my 3D prints, and they failed in completely opposite ways
This title could be clearer and more informative.Try out Clickbait Shieldfor free (5 uses left this month).
A hands-on comparison of Claude Opus 4.8 and GPT-5.5 generating OpenSCAD files for 3D printable parts across three tasks: a phone stand, a circular microphone array enclosure, and a remote cradle. The two models failed in opposite ways — GPT-5.5 consistently did better research (finding real coordinates, downloading CAD files) but produced geometrically broken parts that ignored gravity or had disconnected components. Claude produced cleaner, physically valid geometry but often invented or refused to find real dimensions. Neither model reliably combined good research with correct physical reasoning in a single pass. The write-render-look-fix loop helped, and light human intervention fixed most issues quickly. The conclusion is that LLMs are a useful starting point for parametric CAD work with human oversight, but cannot reliably one-shot a functional design.
