Introduction Large language model (LLM) spending rises as more workflows move from pilot to production. Every request is billed by the token and each new assistant or agent adds requests. Finance asks for a lower bill and engineering has four main ways to deliver it: model routing, caching, prompt trimming and batching.