There is a real problem with the Codex quota: OpenAI detected 3 additional consumption points, and all of them will be reset tomorrow

Compared to Twitter News, AI News, OpenAI has finally found out the specific reason why the Codex quota is dropping so fast. Head Tibo Sottiaux said that the team had discovered 3 additional costs: extremely long sessions with images were inefficient when compressing the context repeatedly; Computer History was overused in some high-consumption scenarios; and even the ability to automatically generate session titles took up a bit more credit than expected. Tibo previously claimed that no overall anomalies were found, and also pointed some affected users to share subscriptions with sub2api. Only yesterday did he further acknowledge that the cache hit rate of some users has indeed deteriorated, which may lead to faster quota consumption. The investigation has now identified a few specific issues with Codex itself. Context compression is to allow Codex to continuously compress old content during extremely long tasks, freeing up context to continue working. OpenAI's official data also confirmed that Codex automatically uses compaction to maintain tasks for a long time. There is an extra waste in this process when there are many images and multiple compression times in a row. Computer History is a recently launched feature that can bring operation records from selected apps and web pages on Mac into ChatGPT and Codex. OpenAI will push the fix on Sunday US time, and also perform a full reset of all paid subscription Codex usage. Tibo said Reset is expected to go live at around 2 p.m. West Coast time, corresponding to around 5 a.m. on August 24, Beijing time. The team also found an unrelated new optimization plan, which they said could significantly improve efficiency and will continue to advance next week.