There is one line in your system prompt that multiplies your bill by 5. datetime. now() The cache reads your prompt left to right and stops at the first difference. A date on line three means the 60,000 tokens of documentation below it get recomputed. Every request. Forever. The same million input tokens: $3.00 if the start of the prompt changed $0.30 if it matched byte for byte Not a discount. Not an enterprise deal. An official line on the pricing page: Input Price (Cache Hit). One agent run, 30 steps: miss → $6.08 hit → $1.22 1,000 runs a day → $1.77M a year apart. On one line of text. Moonshot themselves run 90%+ cache hit on coding traffic. That is their operating norm, not a blog aspiration. If you are at 30%, it is not the provider. 30 seconds, right now: grep -rn "now()\|uuid4()\|random" your_prompt_builder.py Everything it finds above your documents is billed at 10x, every single day. Move it to the bottom, next to the user's question. Tomorrow, check cached_tokens. The cache does not pay you for what you write. It pays you for what you refuse to change.