After 44K views, the comments taught me the cache trick that actually saves your limit
retained text
My last post measured that coming back to a long chat after 1h costs \~20x more. The comments added things I didn't know, so here's the summary: \- Since 2.1.251 the status line JSON has a prompt\_cache field with ttl and expires\_at, so you can show how long until your next message is a cold one. A cache hit resets the timer for free. There are already free status line scripts that show this countdown, search "claude cache statusline" on GitHub. \- Cached reads are billed at 0.1x input. That's why a big [CLAUDE.md](http://CLAUDE.md) barely matters while the chat is warm (I measured +4%) but doubles the first message. \- Subagents don't get the full Claude Code system prompt, but they run on their own cache, so their first request is uncached. A fork shares the parent's cache. \- The cold return grows with context: one commenter worked out \~$1.20 cold vs \~$0.03 warm at 150k on Opus. So the practical rule: if you've been away for more than an hour, don't type "continue". Start a new chat with a short summary instead. Next I'm measuring subagents vs the main chat, and cold returns at bigger context sizes. What else should I test?
document fetched 2026-11-16 00:21 · canonical https://reddit.com/r/ClaudeAI/comments/1xwz486/after_44k_views_the_comments_taught_me_the_cache