How do I reduce AI costs?

AI costs are dominated by LLM API calls. Here are the most effective optimization strategies, ordered by impact: 1) Use smaller models for simple tasks — classify intent with Haiku/GPT-4o-mini ($0.25/M tokens), then route complex queries to Opus/GPT-4o ($15/M tokens). This alone reduces costs 50-70%. 2) Cache frequent queries — if 20% of queries are repeated, caching saves 20% of API costs. Use semantic caching (similar queries return cached results) for even higher hit rates. 3) Optimize prompt length — shorter prompts cost less. Remove redundant instructions, use concise system prompts, compress retrieved context. A 50% prompt reduction = 30-40% cost reduction. 4) Batch non-urgent requests — Anthropic and OpenAI offer 50% discounts on batch API calls. Queue analysis, summarization, and classification tasks for batch processing. 5) Self-host for high volume — at 100,000+ queries/month, running Llama or Mistral on your own GPU breaks even with API costs. Above 500,000 queries, self-hosting is 3-5× cheaper. 6) Streaming and early termination — stream responses and stop generation when you have enough information, rather than waiting for the model to produce maximum tokens. Monitor: track cost per conversation, cost per user, and cost per feature. Set budgets and alerts. AI costs can spike unexpectedly when usage patterns change.