Anthropic's Silent Cache TTL Cut

What Changed

On April 2, 2024, Anthropic silently changed the default cache Time-To-Live (TTL) for Claude Code from 1 hour to 5 minutes. This was not announced in any changelog, and the official documentation still states "up to 1 hour." The change is definitive: analysis of 1,140 sessions shows 100% usage of the ephemeral_1h tier before April 2, a mixed day on April 2, and 100% usage of the ephemeral_5m tier from April 3 onward.

The Impact on Your Workflow and Wallet

The practical effect is severe. For one developer, daily cache busts increased from 39 to 199—a 5.1x multiplier. The associated cost rose from $6.28/day to $15.54/day, projecting a $277.80 monthly increase from this single change.

The shorter TTL creates a compounding problem:

1. Mid-Session Expiry: Cache expires while you're still working, causing Claude to lose confidence in its context.
2. Redundant Reads: Claude re-reads files to re-establish context, padding conversation history.
3. More Expensive Rebuilds: That padded history makes the next cache rebuild even costlier.

A critical, hidden trap is backgrounded tasks. When Claude runs a long tool call, agent, or a command like /loop with an interval over 5 minutes, it suspends the session. If the task takes longer than 5 minutes to return, the cache expires before you see the result. Your next turn pays full price to rebuild the context you just had.

How to Fight Back: Settings and Hooks

You can't change the TTL, but you can change your configuration to minimize the damage.

1. Install the Token Insights Skill

The claude-memory plugin from the gupsammy/Claudest repository provides data and automated hooks.

/plugin marketplace add gupsammy/claudest
/plugin install claude-memory@claudest

Run /get-token-insights in a session. If Claude detects the 5-minute TTL pattern in your data, it will offer to install protective hooks automatically.

2. Configure Protective Hooks (Manual Setup)

Add these hooks to your ~/.claude/settings.json file to get warnings before a cache bust occurs.

{
  "Stop": [
    "plugins/claude-memory/hooks/cache-warn-stop.py",
    "plugins/claude-memory/hooks/cache-warn-3min.sh"
  ],
  "UserPromptSubmit": [
    "plugins/claude-memory/hooks/cache-expiry-warn.py"
  ]
}

• cache-expiry-warn.py: Warns you when a prompt submission is about to trigger a cache bust.
• cache-warn-stop.py & cache-warn-3min.sh: A two-part system that starts a background timer and stops the session before the 5-minute expiry.

3. Adjust Your Global Settings

Cap your context window and enable clearer session management:

{
  "env": {
    "CLAUDE_CODE_DISABLE_1M_CONTEXT": "1",
    "ENABLE_TOOL_SEARCH": "1"
  },
  "showClearContextOnPlanAccept": true
}

• CLAUDE_CODE_DISABLE_1M_CONTEXT": "1": This is the single most impactful setting. It caps your context at 200K tokens instead of 1 million. When a cache bust happens, you rebuild from scratch. A 1M-token rebuild costs ~5x more than a 200K-token rebuild. With busts happening 12x more often than before April 2, this setting is essential for cost control.
• showClearContextOnPlanAccept": true: This adds a button to clear context after accepting a plan. It's useful for separating planning and implementation into cleaner, cheaper sessions.

The Official Stance and Your Action Plan

Boris, the creator of Claude Code, has acknowledged the caching issue on GitHub. The official advice is to start new conversations more frequently to avoid large cache misses, and to be selective with skills/agents to keep context lean.

Your new workflow should be:

1. Monitor: Run /get-token-insights to see your cache bust pattern.
2. Protect: Install the hooks to get warnings before you lose cache.
3. Limit: Cap your context at 200K via the environment variable.
4. Segment: Use showClearContextOnPlanAccept and consciously start new sessions for long-running tasks or after breaks.
5. Feedback: Use Claude Code's /feedback command to report specific instances of costly cache behavior. More data helps Anthropic prioritize a fix.

This isn't just about cost—it's about predictability. A 5-minute TTL turns any pause into a potential budget-buster. Configure your environment now to take back control.