HexLab

Pilot: Prompt caching › Ep 6

Ep 6 · Time to live

Length 1:17

Today's goal: Your cache has a timer.

What you'll learn

  • On Anthropic's API the cache lasts 5 minutes by default, or an hour if your tool asks. Every read restarts the timer.
  • Come back after the timer runs out and even “hi” rewrites everything: about $4.48 on our 559k-token session.
  • A cheap keep-warm ping restarts the timer. In our own measurements it cost $0.09–$0.17, against a $3.66–$6.68 rewrite.

Sources

  • Anthropic's prompt-caching reference (as bundled with Claude Code 2.1.299, 21 Oct 2026): the 5-minute and 1-hour TTL
  • Our own measurements: keep-warm pings on Opus 5.5, 18–19 Oct 2026
  • Claude Code 2.1.299: idle compaction, checked in the installed binary
About our sources →