You only need the frontier model for one single edit
Summary
This article argues that for automated editing tasks, using a frontier model solely for planning can dramatically reduce costs and speed, as the cheaper models can execute the plan. It presents empirical comparisons of different agent configurations (/PLAN and /PREWALK) and shows that reading tokens (context) dominates cost, not just the execution of edits. The piece outlines a pragmatic workflow: start with a frontier model to generate a plan, then hand off to a cheaper model for execution, leveraging structured prompts and todo lists to maintain reliability.