Codex Quota Issue: OpenAI Identifies 3 Additional Usage, Full Reset Tomorrow

Insightful Beating AI Newsflash: OpenAI has finally identified the specific reasons behind Codex's rapid quota depletion. Lead Tibo Sottiaux stated that the team has discovered three additional sources of consumption: inefficiency in compressing context with long sessions containing images, overutilization of Computer History in certain high-consumption scenarios, and even the function of auto-generating session titles exceeded the expected quota.

Prior to this, Tibo first claimed there were no overall anomalies, then directed some affected users to sub2api and shared subscriptions. It was only yesterday that he further admitted that the cache hit rate for some users had indeed worsened, potentially accelerating quota usage. The investigation has now pinpointed several specific issues within Codex itself.

One of these issues, "context compression," aims to continuously compress old content within Codex during extended tasks to free up context for ongoing work. Official OpenAI documentation also confirms that Codex will automatically use compaction to sustain long tasks. However, with many images and multiple consecutive compressions, this process now incurs additional waste. Computer History, on the other hand, is a recently launched feature that allows actions recorded on a Mac's selected apps and webpages to be brought into ChatGPT and Codex.

OpenAI will roll out a fix on Sunday PST and will conduct a full reset of Codex usage for all paying subscribers. Tibo mentioned that the reset is expected to go live around 2 p.m. PDT, corresponding to around 5 a.m. on August 24th in Beijing time.

The team has also discovered an unrelated new optimization solution that is said to significantly enhance efficiency, which will be further pursued next week.