Edited by Sia
Just as Cursor CEO Michael Truell revealed that OpenAI plans to block Cursor users from accessing its models in three months, causing a sudden strain in their relationship, OpenAI Codex has suddenly made a new move.
Just now, these paying users suddenly received a reset of their usage limits—Tibo, head of Codex, announced on Twitter that usage limits will be reset for all paying users of Codex and ChatGPT Work.
This time, it's not simply a matter of giving a bit more.
Tibo said they recently reviewed thousands of user feedback reports and completely overhauled the usage calculation mechanism behind Codex. They found that users’ previous feeling that their credits were depleted too quickly wasn’t just an illusion.
After a series of bugs were fixed, depending on usage patterns, they expect the same Codex credit allowance to now last 10%-50% longer than before. The credit number may remain unchanged, but the actual work that can be accomplished has increased.
This released list of issues is almost a comprehensive guide to “How AI Agents Secretly Burn Tokens.” Some of these bugs may seem minor, but they can consume substantial amounts of tokens very quickly.
One task can consume up to 70% of the weekly quota.
OpenAI listed eight categories of issues that have been identified and resolved.
First is context compression. As an AI programming tool runs for an extended period, the context grows larger, and the system typically compresses historical information to continue functioning efficiently.
The issue was that, when Codex previously performed context compression, it retained old images in the context. As a result, instead of reducing the context size, the unremoved images kept the context large—so large that it could immediately trigger another compression.
This led to a somewhat magical situation: the compression intended to save context ended up consuming even more context. After OpenAI fixed this issue, usage among users who heavily relied on images dropped by approximately 10%.
But that’s not the most extreme. The real quota killer comes from Codex’s goal mechanism.
OpenAI found that in some cases, even though the user-specified /goal had already been completed, the agent did not stop as expected but continued executing subsequent steps.
In another scenario, even though the tool is clearly broken, the model continues to retry repeatedly. The task appears to be finished, but the AI is still working behind the scenes over and over again.
OpenAI stated that in some cases they observed, just this single issue could consume 15% to 70% of a user’s weekly quota. In extreme cases, a single anomalous task could directly use up 70% of the weekly limit. This issue has now been resolved.
There was a background task that nearly ran 15,000 times.
Another issue arises in Memory, the memory system.
The backend Memory Worker in Codex, under certain conditions, inherits certain Stop Hooks. Stop Hooks are intended to control when a task should stop. However, due to this bug, some background tasks may continue running indefinitely because the stop conditions cannot be met.
The users affected by this issue actually account for less than 1%. However, the problem lies in the extreme nature of long-tail cases. OpenAI noted that they even discovered a case where the system checked whether a task could be completed as many as 15,000 times.
For the vast majority of users, this bug may be completely unnoticed. But for those affected, tokens could silently vanish in the background—like a computer’s fan suddenly spinning wildly despite no applications being open.
Subagents also secretly upgrade their configurations.
Another issue characteristic of the Agent era occurs with subagents.
The current Codex does not necessarily rely on a single model to complete tasks. Complex tasks can be broken down and executed collaboratively by multiple sub-Agents. The issue lies here.
OpenAI found that some smaller models, such as Luna, sometimes autonomously select more capable and higher-cost auxiliary models without explicit user request.
Even more surprisingly, even if the main model coordinating the task is not running in /fast mode, it may still require subordinate agents to use /fast.
It’s like the boss flying economy but secretly upgrading a few assistants to business class. The task remains the same, but the underlying resource consumption is no longer equivalent. This issue has now been resolved.
Automated tasks, promised to run once daily, may actually run more often without notice.
Automations have also been found to have issues.
OpenAI stated that some custom scheduling tasks previously might have occurred at a higher actual frequency than the user-set frequency. For example, a user may have intended for an automated task to run periodically, but the system might have awakened the task more frequently than the scheduled interval.
Individually, the cost may seem small, but the hallmark of an Agent is that it can run autonomously without user intervention. Running it once is inexpensive; running it dozens of times may not be. This issue has already been resolved.
Just summarizing what you've done can consume 20% of your weekly limit.
Another common issue comes from Computer History.
For the Agent to understand what it previously did on the computer, the system needs to save, organize, and even summarize past operational history. However, in the previous implementation, Codex might repeatedly summarize historical activities that are already highly overlapping—essentially reorganizing the same segment of work log over and over again.
OpenAI stated that, in some cases, this additional overhead can account for about one-fifth of a user’s total weekly usage—approximately 20%.
There is another similar backend consumption from Rolling Task Summaries. Ordinary conversation turns may also trigger additional backend requests to generate rolling task summaries.
The cost per occurrence isn't significant; OpenAI estimates it increases token usage by about 1%. As before: 1% once isn't much, but it adds up when it happens every time. OpenAI has already disabled this mechanism entirely.
MCP didn’t escape either: the same tool results might be encoded twice.
The final category of issues occurs with MCP tool calls.
OpenAI found that some tools may return results that are double-encoded. Additionally, some tool descriptions were inadvertently truncated, requiring the system to retrieve them again.
Individually, these issues may seem like minor glitches in engineering implementation. But when an Agent calls tools dozens, hundreds, or even more times per day, each repeated transmission represents actual token consumption. Over time, these small costs accumulate and ultimately impact the user’s quota.
An increasingly difficult account to calculate
Looking at this fix as a whole, a clear change is emerging: previously, when using ChatGPT, users essentially viewed a single conversation as a single model invocation.
But with Agent products like Codex, things are completely different.
You entered only one sentence on the interface, but the system may be executing an entire agent workflow—making it increasingly difficult to intuitively understand how much credit has been used.
Some tokens are used by the model to actually write code. Others are used to understand context. And some tokens are generated solely by the backend system to maintain memory, generate summaries, or schedule agents. If any of these steps involve loops, repeated calls, or incorrect scheduling, the user will ultimately experience just one thing: why did my quota disappear without anything being done?
OpenAI clearly recognized this issue. In addition to fixing the aforementioned bug, they stated that architectural adjustments have been made to prevent similar issues from recurring. If similar anomalies reoccur, the team will automatically receive alerts.
More importantly, OpenAI is developing a new usage display feature. In the future, users will be able to see directly within the app exactly where their credits have been spent, without having to guess.
From this perspective, this reset might merely be a superficial benefit. Moreover, it seems OpenAI hasn’t intended to stop here?
This article is from the WeChat public account "Machine Heart" (ID: almosthuman2014), authored by: Focused on Science AI