How do you control the cost of an LLM feature in production?

Route requests by difficulty — small model for mechanical work, flagship only for hard cases. Cache stable prompt prefixes and cache/store repeated results. Cap output tokens and request only needed fields. Use batch APIs for non-interactive jobs (often ~50% cheaper). Retrieve minimal context instead of stuffing documents. And measure tokens per feature in CI so regressions surface immediately rather than in the monthly bill. These commonly cut spend by more than half.