I don't actually disagree, and advocate that for now the thing to do is use both.

You need to use frontier models to understand where the puck is going, but also to use local ones for anything remotely sensitive.