LLM API Cost Structure — Empirical Benchmarks for Caching, Routing, and Distillation in Agent Fleets