OpenAI found 3 additional consumption of Codex quotas

OpenAI finally found out the specific reason why Codex quotas fell so quickly. TiboSottiaux, the person in charge, said that the team has found three additional consumptions: it is inefficient to repeatedly compress the context for long sessions with pictures;ComputerHistory uses too much in some high-consumption scenarios; and even the function of automatically generating session titles is more than expected. Tibo previously claimed that no overall abnormalities were found, and then pointed some affected users to sub2api and subscription sharing. Only yesterday did he further admit that the cache hit rates of some users have indeed deteriorated, which may lead to accelerated credit consumption. Now the investigation has finally identified several specific problems with Codex itself. Among them,"context compression" is to allow Codex to continuously compress old content during ultra-long tasks, freeing up context to continue working. OpenAI official data also confirmed that Codex will automatically use compaction to maintain long-term tasks. When there are many pictures and compressed multiple times in succession, this process now has additional waste. ComputerHistory is a recently launched feature that can bring operation records from selected apps and web pages on the Mac into ChatGPT and Codex. OpenAI will push the fix on Sunday, U.S. time, and make a full Reset for all paid subscriptions of Codex usage. Tibo said Reset is expected to go online around 2 pm West Coast Time, corresponding to around 5 am Beijing time on August 24. The team also found an unrelated new optimization plan that it said could significantly improve efficiency and would continue to advance next week.