My experience is that current models are pretty good at making functional changes without too many more bugs than a human would make, but are still bad at making good high quality changes. I have a degree in software engineering specifically, so perhaps I am overly sensitive to design issues.