LLM API Cost Structure — Per-Token Economics, Caching Strategies, and Model Routing for Agent Fleets