OpenAI now keeps Codex conversation cache indefinitely, adds 12-hour prompt caching to the API

OpenAI will keep the prompt cache for logged-in Codex users with no expiry, so reopening an old conversation hours later still hits the cache.

Per an Oct 20 post on X by Codex lead Tibo Sottiaux, logged-in users across Codex and Work face no penalty for returning to an old conversation or resuming after a few hours, and are "now guaranteed to hit the cache." It is day 16 of Codex's 28-day run of daily releases.

The same day, the OpenAI Developers account announced a 12-hour prompt caching option in the API. The docs say it is available in the Responses API for GPT-5.6 and later. Cache writes cost 2.5x the standard input rate, double the 1.25x for the default 30-minute cache, while reads cost 0.1x for both (0.05x on GPT-6.1 Sol). By the published rates, an agent that sends the same instructions 12 times an hour apart pays roughly a quarter of what the 30-minute cache would cost. OpenAI pointed to persistent agents that reuse the same instructions and tools all day.

Sottiaux's post passed 270,000 views and 4,400 likes within hours. Replies included "Okay THIS one counts as a feature" and "Finally, an actual good change," and one user said it would be the best improvement yet if cache hits mean less limit usage. Another wrote "Two steps forward, one step back" alongside a Codex error reading "Selected model is at capacity."