PRODUCTION SCENARIO
After a prompt-layout change, an engineer wants to confirm that later Claude requests actually reuse cached input.
Which evidence directly establishes a cache read?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: A nonzero cache_read_input_tokens value in the response usage
cache_read_input_tokens counts the tokens actually retrieved from cache on that request, which is the only field that records a hit. cache_creation_input_tokens is the mirror field and records a write, so a nonzero value there means the entry was just created rather than reused. A cache_control block asks for caching, but a prefix below the model's minimum caches nothing and returns no error. And time-to-first-token moves with load and output length, so it cannot separate a hit from a fast miss.