Claude Certified Architect – Foundations Practice Question 89: Prompt Caching Rate Limits
Claude Certified Architect, Foundations · Test 9 · question 9 of 10
A document-QA service sends every request with the same 40,000-token system prompt and tool definitions followed by a user question of about 500 tokens. At peak, the service receives 429 errors citing the input-tokens-per-minute limit on a current Claude model, even though request volume is well under the requests-per-minute limit. The team wants more throughput without requesting a limit increase. What should they do?
Show answer and explanation
Answer:
C is correct. Rate limits are cache-aware: for most Claude models only input_tokens (after the last breakpoint) and cache_creation_input_tokens count toward ITPM, while cache_read_input_tokens do not. With a stable 40,000-token prefix cached, nearly every request counts only the new question against ITPM, so effective throughput rises sharply (and cached reads are billed at a lower rate).
A is wrong because max_tokens governs output, and it does not factor into OTPM calculations or reduce input counted against ITPM. B is wrong because rate limits are shared across all inference_geo values; both draw from the same pool. D is wrong because backoff handles transient 429s but does not raise the sustainable throughput; the same tokens still count against the limit.
Reference: Rate limits - Claude API Docs