florianbuetow Cache Money
Keep the Anthropic prompt cache warm during Claude Code sessions — especially during peak hours when usage limits are tighter — by scheduling lightweight pings tuned to your cache TTL.
Why This Matters
Claude Code sends the full conversation context with every API call. Anthropic caches this prefix server-side and serves subsequent calls from cache at ~10% of the base input price. But the cache expires after a TTL period of inactivity:
| TTL Tier | Duration | Cache Write Cost | Who Gets It |
|---|---|---|---|
| Default | 5 minutes | 1.25x base input | All plans (no ttl field specified) |
| Extended | 1 hour | 2x base input | Max-tier plans (server-side in Claude Code), or explicit ttl: "1h" via API |
Cache reads cost 0.1x base input regardless of TTL tier — that's the 90% saving.
If a session sits idle past its TTL, the next call pays full cache-write price for the entire context — up to 1M tokens. During peak hours (weekdays 5:00 AM – 11:00 AM PT), Anthropic's rolling session limits are consumed faster, making every cache miss doubly expensive: higher rebuild cost plus faster quota burn.
For detailed technical background, consult references/cache-mechanics.md.
Workflow
Step 1: Assess Timing
Determine the current time and day of week in Pacific Time (PT) using:
TZ=America/Los_Angeles date "+%A %H:%M PT"
Classify the current window:
| Condition | Status |
|---|---|
| Weekday, 5:00 AM – 11:00 AM PT | Peak hours active — cache warming strongly recommended |
| Weekday, 4:47 AM – 4:59 AM PT | Peak approaching — pre-warming recommended |
| Weekend or outside peak | Off-peak — cache warming optional, still saves on idle sessions longer than the TTL |
Report the status to the user in one line before proceeding.
Step 2: Detect Cache TTL
Determine which TTL tier is active. Ask the user:
Are you on a Max-tier plan (which enables 1-hour cache TTL in Claude Code), or the default (5-minute cache TTL)?
Use the answer to set the ping interval:
| TTL Tier | Cache Duration | Ping Interval | Safety Margin |
|---|---|---|---|
| Extended (1h) | 60 minutes | 55 minutes | 5 minutes |
| Default (5min) | 5 minutes | 4 minutes | 1 minute |
If the user is unsure, default to the 5-minute TTL (4-minute ping interval) — it's safe for all plans and the overhead is minimal.
Step 3: Start the Cache Ping Loop
Invoke the /loop skill to schedule the recurring ping at the determined interval:
For 1-hour TTL (Max-tier):
Skill tool: skill="loop", args="55m Cache ping. Reply with only: ok"
For 5-minute TTL (default):
Skill tool: skill="loop", args="4m Cache ping. Reply with only: ok"
Each ping triggers one lightweight API call that renews the cached prompt prefix. The response is minimal — just "ok" — so token consumption per ping is negligible.
Step 4: Confirm to User
After starting the loop, report:
- TTL tier: Which cache duration was detected (5-min or 1-hour)
- Interval: The chosen ping interval and why
- Peak window: Whether currently in peak hours
- How to stop: End the session or cancel the loop
- Savings: Cached tokens cost ~90% less than uncached — on a large context this adds up fast
Timing Reference
| Timezone | Peak Start | Peak End |
|---|---|---|
| PT (Pacific) | 5:00 AM | 11:00 AM |
| MT (Mountain) | 6:00 AM | 12:00 PM |
| CT (Central) | 7:00 AM | 1:00 PM |
| ET (Eastern) | 8:00 AM | 2:00 PM |
| GMT / UTC | 1:00 PM | 7:00 PM |
| CET (Central Europe) | 2:00 PM | 8:00 PM |
| IST (India) | 6:30 PM | 12:30 AM |
| JST (Japan) | 10:00 PM | 4:00 AM (+1) |
| AEST (Sydney) | 11:00 PM | 5:00 AM (+1) |
Important Notes
- The cache is per-session. Each active Claude Code conversation has its own cached prefix. The ping must happen within the same session to renew it.
- The ping only works while the session is open. Closing the terminal or ending the session invalidates the cache regardless of the loop.
- Off-peak sessions still benefit if idle periods exceed the TTL — the cache expires based on inactivity, not time of day.
- The first API call after cache expiry is always a full-price write. Every subsequent call within the TTL is a cache read at ~10% cost.