LLM costs can spiral quickly in production. Effective optimization strategies: 1) Semantic caching: cache responses for semantically similar queries (not just exact matches) using embedding similarity threshold ~0.95. GPTCache or custom Redis + embedding implementation. Reduces costs 30-50% for common query patterns. 2) Model routing: use smaller, cheaper models for simple tasks. GPT-4 mini or Claude Haiku for classification, extraction, and simple Q&A ($0.15/1M vs $5/1M). Route to GPT-4o/Claude Sonnet only for complex reasoning. 3) Prompt compression: LLMLingua or selective summarization of context to fit more information in fewer tokens. Can reduce input tokens 50% with minimal quality loss. 4) Batching: batch OpenAI Batch API enables 50% cost reduction for non-realtime tasks (8-24 hour SLA). Good for processing large document sets, bulk classification, overnight analysis jobs. 5) Output length control: set explicit max_tokens, request concise responses explicitly in prompts. Verbose outputs dramatically increase costs. 6) Prompt optimization: systematically test shorter prompts - often 30-40% shorter prompts achieve same quality. Remove verbose instructions, examples that do not help. Monitor cost per feature: attribute API costs to specific product features to understand which are worth their cost. Build cost dashboards in Grafana with per-model, per-feature breakdowns.