In my experience it's not as good as even GPT-3.5 at following instructions (for example formatting a text blob into structured json) which leads it have too much variability in output to be useful.
Have recently been working with it as well and come to the same conclusion. Would be nice to use it since my company is already a GCP shop, but the json parsing is super frustrating. I have noticed using Python's ast.literal_eval helps, since it catches single quotes.