The naive thing is to use one model for everything. We use two: a long-lived Opus orchestrator that plans + dedupes, and short-lived Haiku workers per task.
Why: orchestrators benefit from persistent context (the attack-surface map grows over the scan's lifetime). Workers don't — each gets a small, well-scoped task and exits. Pinning Haiku to workers keeps the cost-per-tool-call near $0.001.