Keep cache forever! OpenAI is a major update, longer than DeepSeek?

AI Application Wind Vane (Public Account: ZhidxcomAI) Author | Bi Weihao Editor | Mo Ying Zhidixi reported on October 21 that just now, ChatGPT and Codex leader Tibo announced that the API service of the OpenAI model will provide a 12-hour cache retention service option, and the Codex and Work login user cache will be retained permanently. This update is also a major upgrade in Tibo's 28-day plan. After many days of resets and small improvements, Tibo finally made a big move. Before releasing the update, Tibo also issued a post stating that it had seen abnormal credit depletion speeds for some users and would refund credit points to approximately 40,000 accounts. Judging from the documentation, the cache retention options on the API side are effective for GPT-5.6 and above models. OpenAI provides default cache retention options of 30 minutes and 12 hours. The permanent cache of Codex and Work has not yet been announced. However, the 12-hour cache does not simply extend the retention time. According to the API documentation of OpenAI, after selecting 12 hours of caching, the cache write price will increase from 1.25 times to 2.5 times the uncached input token price, while the cache read price will remain unchanged. As for the "permanent cache" provided by Codex and Work for logged-in users, its specific implementation still needs to be further confirmed. Whether permanent cache retention refers to the cache state retention policy, the billing arrangement after cache hits, or some kind of rights for users needs to be judged based on the official explanation of this update. However, it is speculated from the mechanism that the permanent cache of Codex and Work is most likely not true cache retention. It may be that the cache is not billed when recalculating, and it is not realistic to permanently retain the cache for users. The OpenAI official API document clearly states that the 12-hour cache setting stipulates the shortest life cycle corresponding to the cache read billing qualification. It does not guarantee that the stored KV tensor can be reused for each request, nor does it guarantee reduced latency. In other words, cache validity, actual cache hits