It struggles with complex concepts and does offer nonsense. (I try to avoid the word hallucinations because that has a meaning that implies agency.)
I'd say the mean acceptance rate of its suggestions is about 50%.
It doesn't do well with new technology. I'm experimenting with bedrock, openai, and langchain. Its suggestions are borderline annoying.
You can ask it to explain a method you don't understand. People say it makes mistakes, but it makes far less mistakes than I do.
It's great for generating boilerplate.
Chatgpt is excellent for explaining the details of things like configuration options.
I still write most of the non-boilerplate code myself.