OpenAI is charging 2.5 times its standard input rate to write reusable prompts under a new 12-hour minimum cache option, giving developers a longer period for repeat API calls.
The company added the option Oct. 20 for GPT-5.6 and later models in its Responses API. A 30-minute minimum remains available at a write rate of 1.25 times the standard input rate. Those multipliers are the total write prices, rather than fees added to the base rate.
On most supported models, reading cached material costs one-tenth of the standard input rate. That lower charge matters when an application sends the same prompt prefix, such as standing instructions or tool definitions, across multiple requests.
The longer minimum gives developers more time to reuse that material at the cached read rate, but they pay more when it is first written. Material that changes frequently can use the 30-minute setting. Whether either choice lowers an application's costs depends on how often it reuses the cached input.
Neither minimum is an expiration deadline. OpenAI says a cache entry may remain available beyond the selected period, so the settings do not specify when an entry will be removed.
In a separate announcement, OpenAI Codex lead Thibault Sottiaux said signed-in Codex and ChatGPT Work users keep cached context indefinitely and are guaranteed a cache hit when they return to an older conversation. His statement concerns those products; the Responses API uses the documented minimum-lifetime settings.
