How a misplaced timestamp broke prompt caching and tripled an AI agent's API bill

Short summary

A team's AI agent API bill tripled because a timestamp placed at the top of every prompt invalidated the entire prompt cache prefix, forcing 18k tokens of static content to be re-processed and fully billed on every call. The fix is simple: place all static content before the cache breakpoint and volatile data (timestamps, session IDs) after it. Frameworks can silently break caching by injecting metadata or reordering tool definitions non-deterministically.

• •A single timestamp at prompt position zero invalidated the entire cache prefix, tripling API costs
• •Static content must go before the cache breakpoint; volatile data after it
• •SDK middleware can silently break caching by injecting metadata or reordering definitions

Generated with AI, which can make mistakes.