OpenAI added a 12-hour cache option to the Responses API on October 20, 2026, giving developers a longer window to reuse unchanged prompt prefixes across separate calls. GEroundtable reports the new TTL extends the usual 30-minute behavior for sessions that may sit idle and resume later.
The tradeoff is price. Writing into the 12-hour cache costs 2.5x the normal input rate, versus 1.25x for the default 30-minute cache. Read hits stay discounted at 0.1x on most supported models and 0.05x on GPT-6.1 Sol.
For builders, that changes the break-even point rather than making caching universally cheaper. Long system prompts, tool instructions, or large reference context may be worth storing for half a day if an agent is likely to wake up and reuse them; one-off requests probably are not.
There is also a separate Codex and Work claim to treat carefully. Codex lead Tibo Sottiaux said logged-in users would keep cache when returning to older conversations or resuming after a few hours. That has been reported, but it is not yet documented in OpenAI API materials, so treat it as a product claim rather than a guaranteed platform contract.