# How do you know your long shared prefix is really being cached? — retained record

> Retained reddit record from reddit:AI_Agents.

Canonical: https://highsignal.app/data/records/98f76074c4b69aef

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:AI_Agents)  published 2026-09-02 14:40 ingested 2026-09-03 02:37

[original source ↗](https://www.reddit.com/r/AI_Agents/comments/1w5c9kl/how_do_you_know_your_long_shared_prefix_is_really/)

## retained text



```
NGL I expected our long shared prefix to cache well. Then we found a request identifier inserted near the top of the prompt, before the stable instructions, tool schemas and retrieved policies. That tiny field changed every call, so the provider reprocessed thousands of shared tokens. Average cost looked acceptable because light accounts dominated the chart, while high volume cohorts paid the repeated prefix cost and waited longer for the first token.

Volatile metadata is pushed back, cache keys are stabilized and tokens are now being attributed based on prompt segments and cohorts. Things have gotten better but I still need some proof that cache hits are responsible for the gains and not the shift in traffic composition.

What’s your method of validation for prefix caching and what metrics do you put more faith in besides token billings and time to first token?
```



document fetched 2026-09-03 00:19 · canonical https://reddit.com/r/AI_Agents/comments/1w5c9kl/how_do_you_know_your_long_shared_prefix_is_really
