OpenAI's API will offer a 12-hour cache-retention option, announced by Tibo, who leads ChatGPT and Codex, according to a 智东西 report. The option applies to GPT-5.6 and later models and sits alongside the default 30-minute retention; choosing 12 hours raises cache-write pricing from 1.25x to 2.5x the uncached input token price while cache-read pricing is unchanged. Tibo also said cache retention for logged-in Codex and Work users would be permanent, but 智东西 notes that no formal announcement had been published and that the implementation and billing treatment remain unclear. OpenAI's API documentation states the 12-hour setting defines the minimum lifetime for cache-read billing eligibility and does not guarantee cache reuse or lower latency, the report said.