Two resets turned into a box generator
What stands out here is not the amount spent, but the outcome: a coding model helped build an application that generates customizable boxes for 3D printing. In the original user discussion, the author says they enabled “Astra” on Ultra in fast mode and used two plan resets in two days on a plan listed at $200.
The primary source is a user report rather than a press release, so this is not a reproducible benchmark. The post provides no information about prompt volume, token usage, code complexity, or the share of manual edits. Two resets cannot be treated as a universal development cost for this kind of project.
Technically, the useful pattern is more convincing than the idea that a model simply draws an STL file. A coding LLM generates parametric Python code, a CAD framework such as build123d creates solid geometry, and the result is exported as STL or OBJ for a slicer.
Boxes are almost an ideal task for this approach. Length, height, wall thickness, lid clearance, dividers, and cutouts can be expressed as parameters instead of manually moving vertices. One script can generate an entire family of parts, which is where a toy generator starts becoming a proper engineering tool.
Printability, however, does not appear automatically. I would first check whether the solid is watertight, whether wall thickness is sufficient, whether mating tolerances work, how overhangs behave, and what happens at extreme parameter values. A polished preview says nothing about whether the lid will snap into place after printing.
Why this is more than a clever little box
The key change is simple: coding models shorten the path from a verbal requirement to an editable physical part. Custom enclosures, mounts, jigs, and rapid prototypes benefit most when a series of variants matters more than one artistic form.
A strong model does not replace the CAD kernel or constraint validation. It speeds up writing the parametric program, while a library builds the actual geometry and the slicer—and ultimately the print—confirms whether it is usable. That is what separates useful automation from impressive-looking file generation.
Two resets in two days sounds expensive, but without iteration data it only signals intense use. The real question is whether an expensive mode produces fewer invalid solids and fewer corrections than a cheaper one, or merely burns through the limit faster. There is no answer yet, but the direction already looks very real.