The two AI models

Settings → AI → Expert model adds a second, stronger model.

The worker does every turn: part searches, writing the design, reading build results and fixing them. Most of the tokens in a board are spent here, on routine work.

The expert is called only when its judgement is worth the price:

When	What it is asked
The requirements are clear	To plan the architecture: which blocks, which parts, which layer count
A build is clean	To review it before ordering, from the build result and the renders
The worker has failed three times in a row	What is going wrong
The worker asks (consult_expert)	Its question

A cheap or local worker with Claude Opus as the expert costs much less than running the expert on every turn, and the plan and review catch the mistakes a smaller model makes. The token count in the header splits usage by model.

Which pairing works best is a question for the benchmark. No recommendation is made until it has been run.