Choose the Subagent Model Against the Task

Finding a filename and explaining a token-refresh failure are different tasks, even if both start by searching the same repository. The second can require careful reasoning about state, time and permissions.

Custom Claude Code subagents can select a model in their configuration. Use that control to test a cheaper option on a well-defined task with a checkable result. The subagent documentation lists the supported choices and inheritance behavior.

Compare valid findings, missed cases, latency and total usage on representative inputs. Include the cost of a stronger reviewer repairing a weak summary. A cheaper request is not necessarily cheaper delivery.

Keep straightforward retrieval deterministic where possible. Reserve additional model capacity for work that benefits from it, based on observed results rather than a blanket rule that reading is easy and architecture needs the most expensive model.