One premium model or several specialized models?

A workflow question about using one strong model versus a role-based portfolio of models.

For AI-heavy work, I keep seeing two different setups. One setup pays for the strongest single model and keeps most context in one place. The other splits work: planning model, execution model, local/private model, review model. The second setup can be cheaper and more flexible, but it needs better handoff packets or the user becomes the integration layer. If the goal is reliable completed work, where does the portfolio approach start paying for its complexity?