At this point, this is categorically false. Your mental model is very out of date.

The results may not be ideal but I can't say I've had a single output so far this year that wasn't usable.

They absolutely 100% produce usable results for large classes of problems. Usually problems on the more "boring" side, but that's most code.

Even local ones are getting "usable." As an experiment I used OpenCode + Qwen coder to code a text based side scroller game in Rust. Left my laptop on and went to bed and in the morning I had a passable demo that looked like the old DOS game ZZT. This was with a 30b model at 4-bit quantization, so pretty tiny by LLM standards.

BTW LLMs seem to do well with Rust. I think it might actually be the guard rails. It lets the LLM iterate against itself. Dynamic languages don't give any compile time feedback, so it's harder to do that. This is something that's 100% the opposite of what I would have expected.