Cache hit guarantee

Prompt caching lowers the price of repeated input tokens, but the hit rate depends on the service provider route rather than on your client. Three OpenAI routing groups therefore carry a published minimum daily cache hit rate:

Group	Daily target
openai-award	65%
openai-stable	75%
openai-premium	85%

If the measured rate for a UTC day stays below the target, the tokens missing from the target are credited back at the price difference between a cache miss and a cache read on those requests.

Make requests eligible

• Send prompt_cache_key on the request. Chat Completions and Responses both accept it. A request without it is not counted.
• Keep a stable prefix for that key, as you would for any prompt cache.
• Use one of the groups above.
• Ordinary text input only. Requests that bill image tokens, audio tokens, or cache writes are excluded, because their pricing cannot share one text-token comparison.

A series — one account, one prompt_cache_key, one model — starts counting only after at least one real cache read has proved the prefix is cacheable, and its eligible input is capped at the largest prefix actually read from cache in that series.

Window and settlement

UTC day window

[00:00, 24:00)

Eligible requests accumulate inside one day.

Calculation starts

00:10 UTC

For the day that has just closed.

Shortfall credited

The tokens missing from the target, at the price difference between a cache miss and a cache read.

Where it lands

Your balance, and the remaining quota of the key that made the requests.

Where to see it

The Cache Hit Guarantee card in Usage shows, per group, the daily target, the measured hit rate, and the amount credited for the selected range.

What is stored

Only a keyed hash of prompt_cache_key is kept with the usage evidence. The raw value is never written to the database, and the guarantee is not advertised at all when that protection is unavailable.