OpenAI to Reset Codex Quotas and Fix Rapid Usage Consumption Bugs
OpenAI is resetting usage quotas for all Codex subscribers and deploying fixes for bugs that caused unexpectedly rapid quota consumption. The issues stem from context compaction during image-based sessions, high usage in the P95 percentile of Computer History, and conversation title generation. This maintenance update directly addresses user complaints about unfair quota depletion, ensuring subscribers get the value they paid for. It highlights the engineering challenges of managing context windows and background tasks efficiently in large language model applications. The quota reset is scheduled for tomorrow around 14:00 PT, and OpenAI has also identified an unrelated optimization method to improve efficiency, which will be developed next week. The bugs specifically impacted long sessions with images requiring multiple context compactions and high-end usage percentiles.
## BACKGROUND
In Large Language Models (LLMs), "context compaction" is a technique used to summarize or prune conversation history so it fits within the model's context window limit. The "P95 percentile" is a statistical metric indicating that 95% of data points fall below this value, meaning the bug heavily impacted the top 5% of users with the highest usage in their computer history.