"Correct" usually means "exactly matches the specification" .... these systems do not do that based on what's been indicated on a bunch of articles, HN discussions, etc. etc. documenting people's collective experiences. It often doesn't one-shot tasks, it requires hand-holding. Oftentimes to a significant degree. Especially if not working on a CRUD webapp with a lot of boilerplate which heavily skew the training data. LLMs return the most likely sequence of individual elements of code, not the most correct code.

So, it's fast if you get lucky and are able to one-shot it, affordable-ish if you get lucky and are able to one-shot it and correct if you get lucky.

And that tends to happen if you're working on a specific set of codebases which are close to the most commonly occurring codebases in the training data set -- i.e. CRUD webapps / heavy boilerplate usage. Most FOSS projects probably don't fit in that camp (i have no data to support this, it's just my gut experience).