ChatGPT Work and Codex Users Get Up to 50% More Usage

OpenAI has restored usage quotas for all paid users of Codex and ChatGPT Work, following the identification and resolution of several issues that were consuming allowances more rapidly than anticipated. The enhancements resulted in a 10% to 50% increase in usage for affected users, contingent on their method of utilizing Codex. The modifications were implemented after the OpenAI team examined thousands of user reports, pinpointing several bugs affecting compaction, memory, automations, subagents, and other elements of Codex.

These bugs were consuming extra usage. One issue pertained to compaction, wherein outdated images occasionally persisted in the context following their cessation of utility, leading to conversations becoming excessively large and triggering further compactions. Resolving this issue yielded a 10% reduction in usage for users heavily reliant on image processing.

Another peculiar memory bug impacted fewer than 1% of users, as background memory workers could continue operating due to inherited 'Stop' hooks that impeded their completion. OpenAI discovered one extreme scenario involving a thread that checked whether it could halt 15,000 times. Errors with 'Goals' could also consume additional allowance, with certain tasks persisting after reaching their predefined stopping point or repeatedly retrying faulty tools.

The company observed instances where custom automations ran more frequently than users had specified. Subagents and 'Computer History' were also found to utilize extra quota. OpenAI's smaller models, such as Luna, were occasionally invoking more potent subagents even when users had not demanded them. An issue arose where a standard-speed orchestrating model would request its subagents to operate in 'fast mode,' resulting in heightened usage.

The outdated 'Computer History' system could repeatedly summarize overlapping activity, consuming up to 20% of a user's weekly usage in some instances. OpenAI has also deactivated rolling task summaries that were prompting superfluous background requests during typical turns, contributing to approximately 1% of token usage. Additionally, certain MCP tool results were being encoded twice, while tool instructions could sometimes be truncated and subsequently fetched again.

OpenAI has confirmed that all these issues have been rectified. Furthermore, OpenAI is introducing enhanced usage tracking and implementing architectural modifications to prevent the recurrence of these problems. The company will automatically notify its teams of any similar issues that may arise in the future. In conjunction with these changes, paid Codex and ChatGPT Work users have received a fresh usage reset, replenishing their allowance while the fixes take effect.