A top-level cache_control field moves the Claude cache breakpoint forward as the conversation grows

The code this removes is short and annoying. On each turn a wrapper walks the message list, strips the cache_control marker off the block that used to be last, and attaches it to the new one. Anthropic's docs describe the replacement in one line: "The cache breakpoint automatically moves to the last cacheable block in each request, so you don't need to update any cache_control markers as the conversation grows" ^[3].

What the server does with that field is prefix matching. It checks whether the prefix up to the breakpoint is already cached from a recent query, uses the cached copy if it finds one, and otherwise processes the full prompt and caches the prefix once the response begins ^[5]. In automatic mode the cached span is everything up to and including the last cacheable block, and reuse happens on later requests carrying the same prefix ^[4].

The same-prefix condition is the thing to test before deleting the wrapper. An agent loop that appends a timestamp, a budget note, or a freshly retrieved document after the history changes the tail of every request. The docs do not say whether the system falls back to a shorter cached prefix when the last block changes, so the second turn's usage output is the only evidence available; the documented example prints response.usage ^[17].

The clock will surprise teams porting from explicit markers. The default lifetime is five minutes and a read refreshes it for no extra cost ^[6], but it runs from the start of the request that writes or reads the entry, not the end of the response, and generation time counts against it ^[7]. The docs give the case directly: a response that takes 4 minutes to stream leaves about 1 minute for a follow-up on the same prefix to start ^[8]. An agent that streams long plans or large diffs misses on turn two while doing nothing wrong. The documented alternative is a 1-hour TTL at 2x the base input token price ^[9], which is twelve times the default window ^[10].

Combining still works. The automatic breakpoint consumes one of the four available breakpoint slots ^[11], which leaves three ^[12] for pinning tool definitions and a system prompt explicitly while the automatic point tracks the growing history.

One caution on price. The footnote multipliers cover hits and refreshes, and only those: 0.025x base input on Claude Fable 5.1 and Claude Mythos 5.1, and 0.1x on every other model ^[13], a quarter of the standard rate for those two ^[14]. Automatic mode writes a new, longer entry on each turn.

The worked example in the docs is a literary-analysis system prompt and a request to analyze the major themes in Pride and Prejudice, on claude-opus-5 ^[16].