A one-hour cache for a five-minute workload
Longer cache lifetimes cost more to write. They only pay back if you actually reuse them.
The extended cache time-to-live is not free: the write costs twice the base input rate rather than 1.25x. You are pre-paying for the privilege of a longer window.
That pre-payment needs two reads to break even instead of one. On a busy endpoint that is trivially satisfied. On a low-traffic internal tool with a request every twenty minutes, the longer window is genuinely the right call — the short cache would have expired.
The mistake in both directions comes from choosing the time-to-live by feel rather than from request spacing. If your median gap between requests is under the shorter window, the cheaper write wins.
1-hour cache write at 2x base rate, on a prompt reused every few seconds.5-minute cache write at 1.25x, refreshed continuously by the traffic itself.Check your own prompt
The analyser checks this pattern along with the other 25, prices each finding against your request volume, and hands back a rewritten prompt. It runs in your browser — nothing is uploaded.
Run the analyserMore on caching
- A cacheable prompt, uncachedReading a cached prefix costs a tenth of sending it again. This prompt qualifies.
- Per-request values above your static contentA cache prefix ends at the first byte that changes. Anything after it pays full price forever.