Prepared by the Nadali editorial team using AI
On 20 October 2026, OpenAI added a 12-hour prompt-caching option in its API. The company says it allows shared instructions, tools and context to be reused for longer when requests are hours apart.
Who it helps
Small-business owners with a custom AI integration or a contractor maintaining one. For example, discuss a workflow that reuses unchanged reply-drafting rules throughout the day. This is an illustrative scenario, not a measured case study. The news concerns the API, not a change to a ChatGPT subscription.
Conditions and limitations
The documentation describes the setting for GPT-5.6 and later models in the Responses API. The default minimum lifetime is 30 minutes. A 12-hour cache write costs 2.5 times the ordinary input-token rate, versus 1.25 times for 30 minutes; these are total rates, not surcharges. Read rates are identical for both lifetimes. Twelve hours is not a data-deletion deadline or a faster-response guarantee.
What to ask your developer
Our editorial recommendation: do not change settings just because the lifetime is longer. Request a short comparison for one workflow:
- Which model and settings are in use now?
- How often do unchanged instructions repeat, and how long are the gaps between requests?
- What do writes, reads and the remaining processing cost together under each option?
- How can the previous settings be restored if the result is unsuitable?
Compare equal workloads and response quality. Do not extend a contractor’s forecast to the whole business without checking it. Nadali has not tested this feature’s savings or speed.
The takeaway: a longer cache is a reason to review one integration’s cost calculation, not proof of a smaller bill. Check your data boundaries separately.
Sources checked on 21 October 2026. Cover: an original AI-generated illustration, not a service interface or evidence of testing.
