OpenAI Identifies 3 Sources of Additional Codex Usage

According to a report by Jinse Finance, OpenAI has finally pinpointed the specific reasons behind the rapid depletion of Codex usage limits. Tibo Sottiaux, the person in charge, stated that the team has identified three sources of additional consumption: inefficiencies in compressing context during repeated operations in ultra-long sessions with images; excessive usage of Computer History in certain high-consumption scenarios; and even the auto-generation of conversation titles consuming more usage than expected. Previously, Tibo had claimed there were no overall anomalies and directed some affected users to sub2api and subscription sharing. However, he further admitted yesterday that the cache hit rate for some users had indeed worsened, potentially accelerating usage depletion. Now, the investigation has finally pinpointed several specific issues within Codex itself. Among these, "context compression" is designed to allow Codex to continuously compress old content during ultra-long tasks, freeing up context space to continue working. OpenAI's official documentation also confirms that Codex automatically uses compaction to sustain long-duration tasks. However, when there are many images and multiple consecutive compressions, the current process results in additional waste. Computer History, a recently launched feature, allows selected app and webpage activity records on Mac to be integrated into ChatGPT and Codex. OpenAI plans to roll out fixes on Sunday, U.S. time, and will perform a complete reset of Codex usage limits for all paid subscriptions. Tibo stated that the reset is expected to go live around 2 PM Pacific Time, corresponding to approximately 5 AM Beijing Time on August 24. The team has also discovered an unrelated new optimization method, which is said to significantly improve efficiency and will continue to be implemented next week.