PRODUCTION SCENARIO
A long customer session on Claude Sonnet 5 caches a 40,000-token system prompt. The app sets the top-level effort to high for the first turn and to low for follow-up turns, after which cache reads drop to zero.
Which change restores cache reads on this conversation?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: Pick one effort level at the start and hold it for the cached session
Top-level effort shapes the rendered prompt, so changing it between requests invalidates the cached prefix, and the docs say to pick a level at the start and keep it constant when the model does not support per-message effort. An effort-only system message really does preserve the cache, but it is a beta limited to Claude Fable 5.1, Mythos 5.1 and Opus 5, and Sonnet 5 rejects it with a 400. A 1-hour TTL and a later breakpoint are both real and both address expiry or placement, neither of which makes an invalidated 40,000-token prefix match again.
Official Source: https://platform.claude.com/docs/en/build-with-claude/effort ↗