A service classifies requests with the same long guide in the morning and afternoon. Caching can reuse computation for their shared beginning, known as the prefix. Its value depends on what repeats and the time between requests.
What your application pays for
On October 20, OpenAI announced a 12-hour caching option for its Responses API. Its guide lists a total write rate of 2.5 times uncached input for «12h», identical read rates across lifetimes and a minimum requested cache lifetime. Check model support, parameters and the data-retention policy that applies to your organization.
What to check in your application
Take a stable fictional guide followed by varied fictional requests. Replay the same schedule with the short lifetime and then with 12 hours, on a compatible model. Record tokens read, written and charged as uncached. If short caching would have yielded zero reuse, savings must cover the extra first write. If requests are already close together, count the reads achieved with short caching too.
Remove secrets and unnecessary identifiers from the shared portion. The requested lifetime helps plan reuse. It does not give a guaranteed deletion deadline. Decide based on the effective cost for your traffic.