PRODUCTION SCENARIO
A user-facing document QA service must answer within seconds and is capped at 2,000,000 input tokens per minute on Claude Sonnet 5. Every request carries the same 150,000-token handbook plus a short question, and the team is about to request a higher limit.
Which change raises effective throughput without a limit increase?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Cache the handbook, since cache reads do not count toward ITPM on this model
On Claude Sonnet 5, as on most Claude models, only uncached input tokens count toward the input-tokens-per-minute limit, so caching the repeated handbook multiplies effective throughput without a limit increase. Each of the other three states something true that does not help: the Batches API does have its own separate limits but cannot answer within seconds, rate limits really are applied per model but each model would then see only part of the handbook, and max_tokens genuinely does not factor into rate limiting, which is exactly why changing it frees no input headroom.
Official Source: https://platform.claude.com/docs/en/api/rate-limits ↗