Claude Code Cache TTL Diagnostic

April 17, 2026 · View on GitHub

Is your Claude Code cache broken? Quick diagnostic for the silent TTL change that increased costs 20-32%

Claude Code Cache TTL Diagnostic

Your Claude Code token costs may have silently increased 20-32%.

In March 2026, the prompt cache TTL was reduced from 1 hour to 5 minutes (#46829). This means pausing for more than 5 minutes causes a full cache miss — and you pay for every token again.

Quick Check: Is Your Cache Working?

Run this in your terminal during a Claude Code session:

# Check your current cache hit rate
claude /cost

Look at the cache_read vs cache_creation ratio:

  • cache_read >> cache_creation = cache is working
  • cache_read ≈ 0 = cache is broken, you're paying full price

3 Things That Break Your Cache

1. Pausing more than 5 minutes

Cache expires after 5 minutes of inactivity. Coffee break = full cache rebuild.

Fix: If you need to think, keep the session alive with a quick /cost check.

2. Any file change triggers git status change

Claude Code includes git status in the system prompt. Edit a file → git status changes → cache breaks.

Fix: export CLAUDE_CODE_DISABLE_GIT_INSTRUCTIONS=1 before starting claude. (Trade-off: no git context)

3. Starting a new session

Every new session starts with zero cache. The initial prompt (~6,500 tokens) is always uncached.

Fix: Don't restart sessions unnecessarily. Use /compact instead of starting fresh.

Hook: Auto-detect cache problems

Add this to your settings.json to warn when cache efficiency drops:

{
  "hooks": {
    "PostToolUse": [{
      "matcher": { "event": "PostToolUse" },
      "hooks": [{
        "type": "command",
        "command": "INPUT=$(cat); TOKENS=$(echo \"$INPUT\" | jq -r '.session_tokens // empty' 2>/dev/null); [ -z \"$TOKENS\" ] && exit 0; echo \"$(date +%H:%M:%S) $TOKENS\" >> /tmp/cc-token-session-$(date +%Y%m%d).log; exit 0"
      }]
    }]
  }
}

Want the full picture?

Diagnose your cache instantly: Cache Health Checker — paste your /cost output, get diagnosis in seconds.

By yurukusa — 800+ hours of autonomous Claude Code operation

Opus 4.7 became the default model on April 16. Token consumption is up to 4x higher, and the safety classifier broke (#49618) — 20+ data loss incidents in 3 days. One-command fix: npx cc-safe-setup --opus47