Not every build task deserves your best model. Boilerplate — a CRUD form, a repetitive component, a config file — is cheap, mechanical work; burning your most expensive model on it is like hiring a senior architect to move boxes. Save the expensive reasoning for decisions that are actually hard to reverse.
The split that matters: fast, cheap models for mechanical, well-specified work; your strongest reasoning model for anything touching architecture, security, or a decision you'd hate to redo. Getting this backwards either wastes budget or under-thinks something that deserved more.
Run the seed on two different models and judge one thing only: did the cheaper model's output need more correction than the pricier one — and was that gap worth the price difference for this specific task?