Prompt caching vs the long LLM conversation: where your input bill actually hides

I kept watching my Claude Code bill climb through long sessions, and most of it was not new work. It was the same conversation getting re-sent every turn. A multi-turn call is stateless, so your client ships the whole history each time: file reads, tool output, old diffs, all of it, and you pay input tokens on that pile again and again. So I built PromptCrunch. It is a drop-in proxy that optimizes the conversation before it reaches the model. Under the hood: it deduplicates superseded code, com...