OpenAI prompt caching retention has been stretched from half an hour to 12 hours. The OpenAI Developers account announced on October 20, 2026 that the Responses API now offers an optional 12-hour cache tier: shared instructions, tool definitions and context that hit the cache again after a gap of several hours are still billed at the cached input rate. Codex lead Tibo Sottiaux added that for signed-in Codex and Work users, cache is now kept indefinitely, so returning to an old conversation after any length of time still guarantees a hit, with no expiry penalty.
The 12-hour tier costs more to write and less to wait
| Tier | Minimum retention | Write price | Read price |
|---|---|---|---|
| Default 30m | 30 minutes | 1.25x the input rate | 0.1x the input rate |
| New 12h | 12 hours | 2.5x the input rate | 0.1x the input rate (0.05x on GPT-6.1 Sol) |
Both figures are minimum retention; the system may keep entries longer. The math of the 12-hour tier is simple: you pay double at write time so the cache survives the hours an agent spends idle, queued, or waiting for a user to reply. For a persistent agent that reuses the same instructions and tools all day, the old setup meant repaying the full uncached price of a long prompt every half hour; with the 12-hour tier, the same context only has to be "kept fresh" at the write rate once. The setting applies only to GPT-5.6 and later models in the Responses API.
The trap is placement: do not let every new turn write at 2.5x
The 12-hour tier is not something to switch on everywhere. If ttl is set at the top level of a request, it also applies to the implicit cache breakpoint, so the new tokens of every conversation turn get written at 2.5x as well, and a long conversation's bill goes up instead of down. The sensible approach is to first understand what prompt caching is and at what granularity it matches, then place an explicit breakpoint at the end of the material reused all day: system instructions, tool documentation and fixed reference documents end there, while the user input that changes daily stays outside the breakpoint. The 2.5x write price only pays for itself when the same material is reused repeatedly across gaps longer than half an hour; one-off questions and short sessions are better off on the default tier.
Codex users do not have to change anything
Beyond the API billing change, signed-in Codex and Work users follow a different rule: their cache is retained indefinitely, so reopening a conversation from days ago neither costs more nor runs slower because of an expired cache. This is day 16 of Codex's 28 days of daily releases. For regular users it removes the hidden cost of "step away for a while and coming back means starting over"; for API developers the signal is clearer still: reusing long context is turning from a matter of luck into a certainty you can buy by the tier.
