OpenAI extends prompt cache in Responses API to 12 hours

On October 20, 2026, OpenAI added the "12h" value to the prompt_cache_options.ttl parameter in the Responses API, allowing the unchanging beginning of a request to be cached for up to 12 hours instead of 30 minutes (write cost is 2.5× the base input token price, read cost is up to 95% cheaper), while the head of Codex announced indefinite caching for Codex and ChatGPT Work users.

Compiled by Sergey KostenchukPublished 2026-10-21Updated 2026-10-21

2026-10-21 Coding OpenAI

💻 OpenAI extends prompt cache to 12 hours

Starting October 20, the prompt_cache_options.ttl parameter in the Responses API accepts "12h" instead of the standard "30m": the unchanging beginning of a request—instructions, tools, context—can be cached for up to 12 hours.

🌍 Previously, a pause longer than 30 minutes would reset the cache and increase the cost of repeated requests—this affected long-lived agents and long sessions. Now, writing to the 12-hour cache costs 2.5× the base input token price (the 30-minute cache costs 1.25×), while reading from the cache is up to 95% cheaper than input.

👤 According to Thibault Sottiaux, head of Codex, logged-in users of Codex and ChatGPT Work have indefinite cache storage: they can return to a session after several hours without refilling the cache. Developers using the API only need to specify prompt_cache_options.ttl: "12h" if instructions remain stable throughout the day.

Source 1: https://platform.openai.com/docs/changelog Source 2: https://platform.openai.com/docs/guides/prompt-caching

Sources