Does a mid-session provider switch result in loading the entire context into the new model, inflating session cost?

I don't think I understand the token/cost implications of this feature