Anthropic priced the session, and idle time is the expensive part

A first-party account of what a Claude Code turn is billed for and what breaks the prompt cache. The figures are Anthropic's own: output tokens are priced "at roughly 5x input", reading from the cache costs 0.1x the input price, and writing into it costs "up to 2x". Every turn re-sends the entire conversation to that point, so the working bill is cache reads on the history, full input price on whatever is new, and output price on the response.

Cache invalidation depends on the request prefix. Requests always go out in the same order, tool definitions, then system prompt, then the conversation with CLAUDE.md at the front, and the cache "has to match from the very start of the request forward". Anything that changes the prefix throws the rest away. /model and /effort are both part of what the cache is keyed on, so switching either mid-conversation prefills the whole conversation again at full price. Fast mode is keyed on too, and the re-prefill happens at fast mode prices. /compact replaces the conversation, so nothing behind the system prompt matches afterwards, which is why the post says compacting is "a lot cheaper before a long break than after one". And /rewind is offered as the cheap alternative when the last few turns went nowhere, because it cuts turns off the end and leaves everything in front of them cached.

Command output over 30,000 characters is written to a file with only a preview and path left in the conversation (BASH_MAX_OUTPUT_LENGTH), so the expensive case is the noisy command that comes in just under the limit and then rides along on every remaining turn. And a /loop "fires as a full turn in the session you set it up in, carrying that whole conversation with it every time", with a cache miss on top if more than an hour has passed.

One limitation, stated in the post itself: the two cost curves illustrating model choice carry the note "Curves are for illustration purposes only. They do not represent real benchmark data." Nothing above is drawn from them.

Why it matters: For the first time, Anthropic has laid out its cost model at enough resolution to expose a lights-out penalty. Every mechanism that makes a session cheap assumes an operator at the keyboard: compact before you step away, rewind rather than compact, keep turns close enough together that the cache is still warm. An hour of idle is a full re-prefill. Unattended operation is structurally more expensive per unit of work than supervised operation under this pricing, and that penalty comes from the billing model rather than from any limit on what the agent can do.