Prompt caching reuses a repeated prompt prefix to cut LLM input costs up to 90% without changing output. Explains how it works, what it costs and when it pays off, given that providers otherwise recompute and charge for the entire prompt on every call.