It's good at refactoring smaller bits of code. The longer the input, the more likely it is to make errors (and you should prefer to start a new chat than continue a long chat for the same reason). It's also pretty good at translating code to other languages (e.g. MySQL->PG, Python->C#), reading OpenAPI json definitions and creating model classes to match, and stuff like that.

Basically, it's pretty good when it doesn't have to generate stuff that requires creating complex logic. If you ask it about tasks, languages, and libraries that it has likely trained a lot on (i.e. the most popular stuff in FOSS software and example repos), it doesn't hallucinate libraries too much. And, GPT4 is a lot better than GPT3.5 at coding tasks. GPT3.5 is pretty bad. GPT4 is a bit better to Copilot as well.