Track token usage, model choice, caching, batching, and routing.

Cost-aware systems use cheaper models for simple tasks, reserve stronger models for hard cases, cache repeated outputs, and limit unnecessary context.