How do you decide if this is a Sonnet, Haiku or Opus kind of question / code task? And the effort?

All is in the title - what's the decision process. I guess it's easy if I want to fine tune an email I'll use the cheap one but then this is also a cheap action so it doesn't matter if you pick an expensive model or a cheap one. But when the task become a bit more complex (like adding a feature to a large code base or writing a proposal plan), the temptation is always to pick the best model with the best effort. How do you guys decide which model to use, and do you really switch each time you open a new thr