Claude Code's /compact costs about 9x more on a cold prompt cache than a warm one

Warm compaction is cheap because the cache sees it as the same conversation. According to the prompt caching page, Claude Code sends a separate request with the same system prompt, tools and history as the session, plus a summarization instruction appended as a final user message ^[10]. The prefix matches, so the request reads it from cache ^[11]. In the lab's warm run that left 2,069 input tokens billed as uncached, about 1.5% of the request ^[19].

The cold case has no match to read. The same page says: "After a break longer than the cache lifetime, there is no cache left to read, so the summarization request reprocesses the full history as uncached input." ^[12] Compaction sits in the "Everything else" bucket, and that bucket's lifetime defaults to five minutes, apart from a few server-controlled helper requests ^[13]. That is shorter than most stand-ups. To get the cold runs, the lab set FORCE_PROMPT_CACHING_5M=1, a variable the docs list for debugging cache behavior ^[14].

Warm is a property of the content as well as the clock. In the post-read run, the files added in the last turn were billed in full ^[4]. We'd treat a compaction straight after a large read as partly cold.

The /clear path costs nothing at the moment of clearing. The costs page says: "When you want a fresh start instead of continuity, /clear costs nothing" ^[9]. The first request after /clear ran 380 to 1,049 tokens larger than the first request after a warm compaction ^[23]. The saving from clearing is the summarization request itself, and the new session starts from the same fixed prompt either way.

These are one lab's figures. They come from Opus 5.5 at list price, one day of runs, and sessions built from real JavaScript files ^[6][7]. For the cold ratio to hold elsewhere, the cache lifetime has to be five minutes and most of the session has to sit in the prefix the summarization request replays ^[10]. A session dominated by fresh tool output would look more like the post-read case.

Automatic compaction adds a size question. On the Anthropic API, models with a native 1M window, including Opus 4.7 and later, compact at about 967K tokens by default ^[16]. The lab's 136k session is about one-seventh of that ^[22], and the figures we have stop there. The window can be set through /autocompact, the autoCompactWindow setting, the --autocompact flag or CLAUDE_CODE_AUTO_COMPACT_WINDOW, and the environment variable takes precedence over the other three ^[15].

In our view the cheaper habit is to compact before leaving a long session, since compacting on return pays the cold price. That holds for Opus 5.5 at list price, a five-minute cache and sessions near 136k tokens.