What changed
OpenAI’s October 20 API changelog adds a selectable 12-hour minimum prompt-cache lifetime. The guide lists it for GPT-5.6 and later in the Responses API. This is a developer feature for reusing an eligible input prefix, such as shared instructions or reference material, rather than a ChatGPT subscription setting.
The guide puts the default 30-minute write at 1.25 times the ordinary uncached input-token rate, versus 2.5 times for a 12-hour write. These are total write rates, not extra surcharges. Read rates do not change with the selected lifetime. The eligible prefix must contain at least 1,024 visible input tokens.
A longer window is not a deletion deadline
The 12 hours describe minimum eligibility for cache-read pricing, not a promise that data is deleted then. Maximum retention policy is unchanged. OpenAI also says the setting does not guarantee reused stored computation or faster replies. Upgrading a valid 30-minute prefix to 12 hours incurs a new write charge without credit for the earlier write.
Our take
For API tinkerers and experienced technicians, start with the shape of the workload: how much of the prompt stays identical, and how long passes before it is reused? A daily reference document and a frequently changing conversation may call for different choices.
Compare total writes and reads over representative traffic before predicting savings. A lower read rate cannot compensate for reuse that never happens. Keep content order stable where that makes sense, and avoid adding filler simply to reach a cache threshold.
We have not called the API or measured billing or speed. Treat the documentation as a reason to evaluate your application, then make a change only if the observed results justify it. The vendor’s dated entry establishes October 20; its exact posting hour was not established.