AI Briefing
KO

Claude cache 5 minutes

·2026.04.23 20:41

Key point

Evidence has emerged that Anthropic lowered Claude Code's default prompt cache TTL from 1 hour to 5 minutes.

1 / 2

Details

Evidence has emerged that Anthropic lowered Claude Code's default prompt cache TTL from 1 hour to 5 minutes.

  • In the JSONL logs under ~/.claude/projects/, usage.cache_creation is split into ephemeral_5m_input_tokens and ephemeral_1h_input_tokens.
  • One user analyzed 1,140 sessions and found that from March 1 to April 1, all sessions used the 1h cache; on April 2 there was a mix; and from April 3 onward, all sessions used the 5m cache.
  • Anthropic explained in a GitHub issue and social media replies that the 5-minute cache is the default, and claimed that overall cost is lower because there are many one-shot requests.
  • However, for long-context agentic work, cache busts surged sharply: for an example account, daily busts went from 39 to 199, and cost went from $6.28 to $15.54.
  • By the article's estimate, this change alone could add $277.80 to monthly costs, with an even bigger impact felt on 1M context and background tasks.

Checking the cache_creation entries in session logs lets you directly distinguish whether the 5m or 1h cache is being used.

This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.

Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.