Prompt Caching Cut My Claude Bill by 80%: The Mistakes That Were Costing Me

Short summary

A developer shares three silent bugs that prevented Claude prompt caching from working: dynamic dates in the system prompt, unsorted JSON keys changing byte order, and per-user tool lists at position 0. After fixing prefix stability and adding a single cache_control marker, cache reads went from zero to consistent, cutting input token costs by roughly 80%. The post includes concrete code examples and a debugging checklist.

• •Dynamic dates, unsorted JSON keys, and per-user tool lists silently invalidated the cache prefix
• •Moving volatile content after the cached prefix and sorting keys fixed cache hit rate from zero to consistent
• •Cache reads cost ~0.1x base input price; input bill dropped ~80% after fixes

Generated with AI, which can make mistakes.