measure twice, cut once. I do this too, often using heavier models with large context windows for design and handing off implementation to cheaper models.