Coming back to an old Codex thread? Codex now keeps your prompt cache with no time limit, so resuming costs no cache penalty

#What changed

On Oct 20, 2026 (PT), Tibo (@thsottiaux), who leads Codex at OpenAI, said that for logged-in users of Codex and ChatGPT Work, OpenAI now keeps the prompt cache with no expiry. Coming back to an old conversation, or resuming after a few hours, no longer drops the cache. He says those turns are now guaranteed to hit it. He posted this next to the API's new 12-hour cache option.

#Who is affected

Anyone signed in to Codex or ChatGPT Work with a ChatGPT account who keeps long threads and returns to them later. If you call the OpenAI API with an API key, the API's own cache rules still apply, including the 12-hour option.

#What to do now

• You no longer need to start a fresh thread just to avoid paying for a cold cache when you come back to old work.
• The cache only matches an identical prefix. Keep AGENTS.md and other stable instructions at the top, and keep things that change every turn, like the date or git status, out of the opening, as @rthiago pointed out.

#What is not confirmed

OpenAI has not published documentation for this yet. Tibo did not say which plans it covers, how cached tokens count against Codex usage limits, or whether threads started before the change are included.

Sources:

• https://x.com/thsottiaux/status/2112712984644784534
• https://x.com/rthiago/status/2112718825791582560