OpenAI is resetting usage limits for all paid Codex and ChatGPT Work users after identifying and fixing a series of bugs and architectural issues that were consuming tokens unnecessarily.
According to the company, users should now be able to get 10% to 50% more usage from their existing limits, depending on how they use Codex.
The changes follow an intensive investigation that involved reviewing thousands of user reports and deploying fixes across several parts of the Codex system. OpenAI said its teams examined the platform “with a fine comb,” uncovering a number of longstanding issues that were quietly increasing usage.
Compaction Was Retaining Old Images
One of the problems involved context compaction. OpenAI found that older images were sometimes being retained when conversations were compacted, causing contexts to remain unnecessarily large and, in some cases, triggering additional compaction.
After the fix, users who rely heavily on images saw usage fall by roughly 10%.
Background Memory Tasks Could Keep Running
OpenAI also identified an issue with background memory workers. These processes could inherit Stop hooks and continue running even when the hook prevented them from completing normally.
Although the issue affected fewer than 1% of users, OpenAI said the impact could be significant for those affected. In one extreme example, a single thread reportedly checked whether it could stop as many as 15,000 times.
The issue has now been fixed.
Goals Could Continue Beyond Their Intended Limits
Another significant source of wasted usage involved Goals. In certain situations, a /goal could technically finish but continue operating beyond its intended stopping condition.
The system could also repeatedly retry malfunctioning tools instead of stopping. OpenAI said individual incidents were found to consume between 15% and 70% of a user’s weekly allowance.
Those behaviors have now been addressed.
Automation Schedules Were Sometimes Running Too Frequently
OpenAI discovered that some custom automation schedules could execute more frequently than users had configured.
The scheduling issue has since been corrected to ensure automated tasks follow their intended frequency.
Subagents Were Sometimes More Capable Than Requested
The company also found problems involving subagents. Smaller models, including Luna, could sometimes select more capable helper models even when users had not explicitly requested them.
Similarly, an orchestrating model that was not operating in /fast mode could sometimes ask its subagents to use /fast.
OpenAI has fixed both behaviors to better align subagent selection and execution with the user’s configuration.
Computer History Was Repeatedly Summarizing Activity
The older Computer History implementation could sometimes repeatedly summarize overlapping activity. In certain cases, OpenAI said this behavior consumed as much as 20% of a user’s weekly allowance.
The underlying implementation has now been corrected.
Rolling Task Summaries Added Extra Requests
Even ordinary interactions could trigger additional background requests through rolling task summaries. OpenAI estimated that these requests added approximately 1% to overall token usage.
While relatively small on an individual basis, the company noted that the additional consumption could accumulate over time. The behavior has now been disabled.
MCP Results Were Sometimes Processed Twice
OpenAI also uncovered issues involving Model Context Protocol (MCP). Some tool results were being encoded twice, while certain tool instructions could be truncated and subsequently fetched again.
These problems have also been fixed.
Usage Limits Reset for Paid Users
Beyond fixing the individual bugs, OpenAI said it has made broader architectural changes designed to prevent similar problems from returning. The company plans to monitor these systems more aggressively, with teams receiving alerts if comparable issues emerge again.
OpenAI is also working on giving users greater visibility into where their usage is being consumed, reducing the need to estimate or guess which activities are responsible for token consumption.
For now, the company is resetting usage limits for all paid Codex and ChatGPT Work users.
With the fixes in place, the expected improvement is substantial: depending on individual usage patterns, users may be able to get 10% to 50% more work from their existing allowances.
The reset comes as OpenAI continues to refine Codex and its broader AI tooling, with a growing emphasis not only on model capability but also on making usage more efficient, predictable, and transparent.
