The arguable caveat is Sonnet may run faster (although this isn't known for sure, due to more tokens being used for the same task), so you can potentially get more done in a synchronous iterative workflow

I don't really believe this however, because so much time is spent fixing up after models, that a slower but more intelligent model is a net time saver in my experience.