OpenAI Codex keeps your chat cache 12 hours
OpenAI made its prompt cache last 12 hours on the 20th of October 2026, day 16 of its 28-day Codex pledge. Coming back to a long chat after a break now costs far less.
OpenAI's Codex keeps chats for 12 hoursOpenAI's Codex now keeps the memory of a chat ready for 12 hours, the company said on the 20th of October 2026. Codex is OpenAI's AI helper that writes computer code.
The change was day 16 of a 28-day pledge, OpenAI's forum page says. Each day brings one clear upgrade or a reset of usage limits, the cap on use.
What a cache is, in plain words
A cache is a short-term memory. The AI keeps the start of a long chat ready, so it skips reading it again at full price.
In OpenAI's API, the paid door apps use, that memory faded after 30 minutes, its guide says.
Days 12 to 17: more use, an edit button, a longer memory and 3 resets
On day 12, the 16th of October, Codex and ChatGPT Work gave more use for the same limits, OpenAI's Tibo Sottiaux wrote on X. They now share memory between helper agents and keep it longer.
Days 13, 15 and 17 were resets, his posts and OpenAI's forum show. On day 14, the 18th in US time, desktop app users could edit earlier messages in a chat, he wrote.
Below we break down what the 12-hour cache costs and saves
These prices are for apps that pay per use through the API. Codex users on ChatGPT plans get usage limits.
On OpenAI's GPT-6.1 Sol, filling a 12-hour cache costs 5 dollars per million tokens, against 2.50 for 30 minutes, its price page shows. Reading from it costs 0.10, against 2 for normal reading.
Take a chat start of 100,000 tokens, small pieces of words, used 6 times 2 hours apart. The 30-minute cache fills again each time: 1.50 dollars, by our arithmetic.
The 12-hour cache fills once and is read 5 times: 0.55 dollars, 63 percent less.
It pays off from 2 returns after a long break.
What is confirmed and what is not
The 12-hour option is on OpenAI's changelog of the 20th of October. For signed-in Codex and Work users, Sottiaux wrote: "we keep cache forever".
OpenAI's forum list calls it a 12-hour cache. Its guides name no endless cache.
Key facts
12-hour cache in OpenAI's API from the 20th of October 2026.
Usage resets on days 13, 15 and 17.
Six uses 2 hours apart: 0.55 dollars with the 12-hour cache, against 1.50, by our arithmetic.