Cut LLM inference costs with architecture-first techniques: model routing, context optimization, prompt caching, semantic caching, batching, and observability.