To check I'm not talking nonsense I asked GPT-4.1 to replace a 30 line Python file I have to access a public API and store the returned data in a database. Took 10 minutes to write and test.
Given no context, it produces code that looks fine. It doesn't even vaguely work as it gets the API syntax (which is publicly documented) very wrong. Fixing it requires replacing many lines of code (which are now 56, not 30).
Given my existing script, it rehashes it to 76 lines without adding value and introducing a few unwanted side-effects. When told what the problems are, the code works, because the meat of it is the existing code. When asked to improve some timestring parsing that was written in a hurry it happily introduces bugs whilst claiming it has been improved.
It will get you 80% of the way there, but so will copy-pasting the documented example from your provider's readme. The rest is from you if you actually want it to work, at which point you might as well just write the code.