I can feel the temptation to point the best model at everything just because it is available.
That is probably wasteful. The premium model should not become expensive bulk labor. Its job should be to harden the machine: schema, compiler, queue contract, brittle assumptions, adversarial review, and the first golden examples that cheaper agents can repeat.
The goal is not to spend intelligence. The goal is to make the system smarter after the expensive window closes.