OpenAI Adds 12-Hour Minimum Prompt Cache for GPT-5.6 and Later

OpenAI has introduced a 12-hour minimum prompt-cache lifetime for GPT-5.6 and later in its Responses API, giving developers a way to reuse shared context across requests made hours apart.

The option, introduced Oct. 20, joins a 30-minute minimum lifetime. Neither setting specifies an exact expiration time: cache entries may remain available longer. Developers can choose the longer minimum for instructions, tools and other context reused throughout the day, while using the shorter minimum for conversation content that changes more often.

Keeping that context cached has an upfront cost. A write with the 12-hour minimum is billed at 2.5 times the standard input-token rate, compared with 1.25 times for a 30-minute write. Subsequent reads cost one-tenth of the standard input rate on most supported models. The pricing benefit depends on how often an application reuses the cached material.

A separate change affects signed-in Codex and ChatGPT Work users: their cached context is retained indefinitely. Returning to an older conversation or resuming one after several hours produces a cache hit. That product retention does not change the API’s minimum-lifetime settings.