Mixing models: plan with one, build with another

Plan with one model, build with another

Plan mode - Cursor’s posture for consequential changes, where it asks clarifying questions, researches read-only, and writes an editable markdown plan you approve before anything is built (modes chapter) - predates the trick this page is about. What shipped in Cursor 2.0 is the split itself: you can create your plan with one model and build it with another. Cursor’s 2.0 changelog frames it directly - “Create your plan with one model and build the plan with another,” with the build running in the foreground or background, and parallel agents able to produce multiple plans for you to review (changelog). Point a frontier reasoning model at the design, where a bad call is most expensive, then hand the approved, written-down plan to a faster, cheaper model to execute.

Read as a cost strategy, the workflow is a quality trick and a spending move at once. A stronger planner runs for the design; then a lower-cost model does the bulk typing against a plan that’s already sitting in a markdown file. The second model doesn’t need the first model’s reasoning context - it needs the plan, which is right there. What the split saves depends on your plan and current routing, so check the usage dashboard (the Spending tab in your dashboard; usage and limits, as of September 2026) rather than assuming a fixed saving. That’s spend-where-it-pays as a two-model loop: heavy reasoning where the fork is, efficient throughput where it’s just execution.

Try it

Switch to Plan mode with the strongest model you have access to, and describe a change with a real design fork in it. When the plan lands, switch the model picker to a cheaper model or Auto (Cursor’s automatic model choice), and build from the plan file. Watch the usage dashboard: the expensive engine ran once, on the design.

None of this needs a memorized roster. The chapter’s judgment fits in one sentence: let Auto route the everyday, pin something stronger only when you can name the turn that earns it, and when a task splits cleanly, give the design and the typing different engines. Max Mode, where you still meet it, is the same call one level down - more context, bought on purpose, on the legacy plans that still offer it (as of September 2026; Max Mode docs).

Which model answers is one dial. What the model knows about your project before it answers is another - and that one you control by writing it down once, in a file the agent reads on every turn. Next: writing rules down once.