OpenAI has reset Codex usage limits for paid users after fixing several technical problems that were causing some customers to consume their allowances faster than expected. The company says users should now receive between 10% and 50% more effective usage, depending on how they use Codex.
The changes follow an internal review of thousands of user reports. OpenAI identified several issues affecting features such as context compaction, memory, automations, goals, subagents and computer history. According to the company, these problems sometimes caused background processes or repeated tasks to consume usage unnecessarily.
Bugs Were Consuming Extra Quota
One of the problems involved context compaction. Codex sometimes retained old images in a conversation even after they were no longer necessary. This could cause the context to become unusually large and trigger another compaction process almost immediately.
OpenAI said fixing this problem reduced usage by around 10% for users who frequently worked with images.
Another issue involved memory workers. The problem affected fewer than 1% of users, but in extreme cases it could cause background processes to continue running when they should have stopped.
OpenAI discovered one case where a thread checked whether it could stop approximately 15,000 times. Although this was an unusual example, it demonstrated how an apparently small software problem could result in unnecessary usage.
Goals Could Consume Large Amounts
OpenAI also found problems with the Goals feature. Some tasks continued operating after reaching their intended stopping point, while others repeatedly attempted to use tools that were no longer working.
These repeated operations could have a significant effect on Codex usage limits. OpenAI reported examples in which such tasks consumed between 15% and 70% of a user’s weekly allowance.
Custom Automations created another potential problem. In some situations, automated tasks could run more frequently than the schedule configured by the user.
This meant users could lose part of their weekly allowance without realizing that a background automation was responsible.
Subagents Created Unexpected Usage
OpenAI’s investigation also uncovered problems involving subagents, which are smaller or specialized AI processes that can be used to perform parts of a larger task.
In some cases, smaller models could unexpectedly call more powerful subagents even when the user had not specifically requested them. Because more capable models can require greater computing resources, these calls could increase consumption.
Another issue involved the /fast mode. An orchestrating model operating at normal speed could sometimes instruct its subagents to use the faster mode, creating additional usage that users may not have expected.
These issues made it harder for customers to predict how quickly they would use their available quota.
Computer History Added More Consumption
OpenAI also identified a problem with the older Computer History system.
The system could repeatedly summarize overlapping activity, resulting in unnecessary processing. In some cases, this behavior consumed as much as 20% of a user’s weekly allowance.
The company has now disabled rolling task summaries that were generating unnecessary background requests during ordinary interactions. OpenAI said these requests accounted for approximately 1% of overall token usage.
Additional technical problems were also discovered. Some MCP tool results were being encoded twice, while tool instructions could occasionally be truncated and then retrieved again. Both behaviors could create additional processing and contribute to higher consumption.
OpenAI Resets User Allowances
Following the fixes, OpenAI has issued a fresh reset for paid Codex and ChatGPT Work users. This means eligible users receive their available allowance again rather than simply having to wait for their normal usage period to end.
The reset is intended to compensate users while the company implements the changes.
OpenAI says the fixes should result in more efficient usage overall. Depending on individual workloads and the features they use, customers could see between 10% and 50% more effective capacity.
However, the exact improvement will vary considerably. Someone who rarely uses images, automations or subagents may see a smaller difference than a user whose workflow depends heavily on those features.
Better Usage Tracking Is Coming
OpenAI is also working on improved usage tracking. The company wants users to be able to see more clearly where their allowance is being consumed.
The planned visibility should help customers identify which tasks, models and features are responsible for the greatest amount of usage.
This could become particularly useful for developers and professionals who use Codex for long-running coding projects. Instead of simply seeing that their allowance has decreased, users could eventually understand what caused the consumption.
OpenAI has also made architectural changes intended to prevent similar problems from returning. The company says its systems will automatically alert internal teams if comparable issues are detected in the future.
What the Changes Mean for Users
The latest fixes are important because usage limits can directly affect how developers work with AI coding tools. Unexpected background activity can reduce an allowance even when users are not deliberately running additional tasks.
By correcting these problems and resetting Codex usage limits, OpenAI is giving paid customers more predictable access to the service.
The company is also taking steps to improve transparency around consumption. Better tracking should allow users to manage their workloads more effectively and avoid unexpectedly exhausting their weekly allowance.
Paid users can take advantage of the refreshed allowance while OpenAI continues monitoring the system. The longer-term goal is not simply to provide more usage, but to ensure that the usage customers receive is spent on the tasks they actually intend to run.
