Models & credits

You came out of the Permissions chapter having tuned one dial: how much Copilot is allowed to do on its own before it checks with you. You kept it tight on shared-lib, loose on orders-service — same engineer, opposite settings, set by blast radius.

This chapter is the other dial. It doesn’t change how much Copilot may act; it changes how hard Copilot thinks. Some tasks want a deeper, more deliberate reasoning pass — the kind of slow, careful read the shared-lib refactor demanded. Others are routine enough that reaching for the heavyweight option just spends more for no better answer. And it does cost more: in Copilot, more reasoning and more context aren’t free, so this dial has a meter attached.

The logic is the one you already know from the action dial — spend where it pays. The whole skill is recognizing which kind of task is in front of you before you turn the knob.

Two lessons:

• Choosing a model per request — the model picker, how selection works per request, and when to reach for more reasoning versus less.
• Spending credits where it pays — what the heavier settings actually consume, and how that’s shaped by the plan your company gives you.

Start with the picker — the control itself, and the judgment behind it. Choosing a model per request.