Isn’t this obvious? When you have a task you think is hard. You give it to a cleverer model. When a task is straight forward you give it to an older one.
If I have a straight forward task, I give it to an LLM.
If I have a task I think is hard, I plan how I will tackle it, and then handle it myself in a series of steps.
LLM usage has become an end in itself in your development process.