OpenAI lets agents cache fixed context for twelve hours: costlier to write, cheaper to reuse

OpenAI now lets developers keep shared instructions, tools and background material cached for twelve hours in its API, instead of the default 30 minutes. The option is aimed at AI agents that work throughout the day and reuse the same context hours apart. For organisations that pay OpenAI per token, it is a new choice that can lower the bill for always-on agents, but only where the same text is actually reused.

OpenAI's developer account on X announced the change on 20 October 2026 at 19:56 Oslo time. The same day it appears in OpenAI's API changelog, and the prompt caching documentation has been updated with prices and rules.

What is new

When a language model reads a long prompt, it computes intermediate results for each part of the text. With prompt caching, OpenAI keeps those intermediate results. The next request that starts with exactly the same text can reuse them instead of computing everything again.

On the newest models, the default minimum lifetime for such a cache entry is 30 minutes. Developers can now choose at least twelve hours instead. According to the documentation, the option applies to GPT-5.6 and later models, and only in OpenAI's Responses API. Developers can set twelve hours for a whole request or only for selected parts, such as fixed instructions and reference material, while conversation content that changes often keeps 30 minutes.

OpenAI uses its own agent, dots, as an example in the announcement. Agents like this run all day and fetch the same background again and again, often with long gaps in between.

More expensive to store, cheap to reuse

The prices are listed in the documentation as multiples of the model's ordinary input-token price:

• Writing text to the cache for 30 minutes costs 1.25 times the ordinary price.
• Writing text to the cache for twelve hours costs 2.5 times the ordinary price.
• Reusing cached text costs 0.1 times the ordinary price on most of the new models, and 0.05 times on GPT-6.1 Sol, regardless of the chosen lifetime.

That makes twelve-hour caching more expensive to start, but cheaper over time if the text is reused many times. By our own calculation from these rates, the option only pays off when the same text is reused at least twice after it was written. If it is reused only once more, it costs more than not caching at all.

OpenAI also writes that the lifetime is a minimum. The company may keep entries longer, and the setting does not change the upper limit in OpenAI's data retention policy. The announcement does not say whether the option is available to customers with a Zero Data Retention agreement.

What the announcement does not say

OpenAI does not state how much a typical customer saves. The documentation stresses that a cached prefix does not guarantee reuse on every request, because the request has to reach a machine that holds the cached entry. It is also not stated whether twelve hours will come to older models or to the Chat Completions interface.

What leaders should note

• Organisations running always-on AI agents on OpenAI should ask their team to pull actual cache usage from OpenAI's dashboard before turning on twelve hours. The gain depends on how often the same text is reused.
• The twelve-hour option can make agents that work through the business day cheaper to run, but it doubles the price of writing the text the first time. Used wrongly, it raises the bill.
• Data protection officers should know that cached intermediate results derived from customer text can remain with OpenAI for at least twelve hours when the option is used. Check how this fits the organisation's data processing agreement before using it on personal data.

Sources and media

• OpenAI Developers on X: 12-hour prompt caching, post published 20.10.2026 at 19:56 Oslo time
• OpenAI: API changelog, entry dated 20.10.2026
• OpenAI: Prompt caching documentation
• Sources were checked on 20.10.2026.
• Illustration: AI-generated with Grok Bot / hogby.ai