Upstream feature request: intermediate cache TTL tiers (15m, 30m)

[cnighswonger]

We filed a feature request on the Claude Code repo for intermediate prompt cache TTL tiers:

anthropics/claude-code#47425

The Anthropic API currently supports only two cache TTL values: 5 minutes and 1 hour. Neither is ideal for interactive use:

• 5m expires during normal human pauses — reading output, thinking, switching tabs. A 6-minute pause triggers a full cache rebuild (100K-900K tokens of cache creation).
• 1h covers interactive work but costs 2x base input on every write. Overkill for short or bursty sessions.

A 15m or 30m tier would cover the gap where most interactive pauses actually land, at lower infrastructure cost than the 1h tier Anthropic already maintains.

From our telemetry (1,610 calls over 61 hours): median inter-turn gap for interactive work is 2-8 minutes — well within a 15m window, frequently outside a 5m window.

This is a win-win: users get better coverage for natural pauses, Anthropic serves cheaper-to-maintain cache entries. The intermediate tiers are less resource-intensive than the 1h tier that already exists.

If this affects you, a thumbs-up on the upstream issue helps surface it to the Claude Code team.