Summary of Key Points
Recently, Codex reset users' quotas for two consecutive days (June 29-30) due to a bug causing abnormal quota consumption. Many users ran out of their entire quota after sending just one message, and the quota even decreased without any user interaction. After an official investigation, it was found that the issue wasn't caused by a single bug but by a combination of several problems: overly frequent automatic code reviews, improper task decomposition, repeated retries, and incorrect statistics. The team adopted an aggressive collaborative approach called "regional joint defense" to quickly iterate on solutions, but this also led to a high frequency of bugs. Although the company provided compensations in the form of "forced resets" and "reset cards," users prefer a permanent fix. However, cutting-edge AI products like Codex may remain in a "semi-finished" state for an extended period.
I. The Abnormal Quota Usage: Not One Bug, but a Cumulative Problem
The sudden increase in quota usage was not due to a single malfunction; rather, several minor issues exacerbated under specific circumstances:
1. Overly Active Automatic Code Reviews: The feature designed to check code quality was activated too proactively. Even without user interaction, the background would start analyzing code and consume quotas silently.
2. Excessive Task Decomposition: A single user request could be split into multiple sub-tasks (understanding, reviewing, verifying, etc.). Each time the front-end was triggered, a series of background processes were executed, rapidly depleting the quota.
3. Repetitive Failed Tasks: After a task failed, the system would attempt to restart it repeatedly, consuming tokens even if no result was produced.
4. Inaccurate Statistics: For example, the traffic from background reviews was mistakenly counted towards GPT-5.4 usage, or unfinished tasks were included in the quota consumption records, leading to discrepancies between the displayed and actual quota usage.
The company later rolled back the changes, fixed the issues with duplicate generation and retries, and issued reset quotas to users again.
II. The Mystery of Hidden Quota Consumption
Users were puzzled by the continuous decrease in their quotas despite no obvious actions. There are three main reasons for this:
1. Background Tasks Consuming Quotas: In addition to automatic code reviews, there's a "memory preview" feature that continuously captures recent screen updates to make conversations smoother. This process consumes tokens even when users are not using Codex (can be disabled in settings).
2. **“Ghost” Quota Usage”: Failed tasks still consume tokens, even if no results are provided. It's like ordering food that never gets delivered but the payment is still charged.
3. Inaccurate Reporting: Different models' token usage was muddled together, and the quota distribution over time (e.g., 2% increase every 5 hours vs. 1% per week) was incorrect, leading users to believe they hadn't exceeded their limits when they actually had.
III. Aggressive Team Collaboration: Fast Updates, but Also More Bugs
The Codex team uses a "regional joint defense" approach to develop the product. In traditional companies, this would involve a PM writing requirements, designers creating interfaces, and engineers writing code in a sequential process. However, at Codex, anyone who is close to solving a problem can start working on a solution immediately. This allows for rapid updates and quick experimentation, but it also leads to bugs being introduced as the product evolves. For instance, billing issues were detected at the end of 2025, and rewriting the underlying system didn't completely resolve them; this time, new problems have emerged.
IV. Evolving Compensation Methods: From “Forced Resets” to “Reset Cards”
The company has changed how it compensates for quota issues:
- Forced Resets: The company directly resets users' quotas, but if the weekly quota is about to refresh naturally, such a reset would be unnecessary (like getting a bonus before your next payment).
- Reset Cards: The company stores reset quotas in users' accounts, allowing them to decide when to use them to avoid waste.
Users are still dissatisfied, as they want a permanent fix rather than temporary compensations for each issue.
V. The Norm of AI Products: Semi-Finished State Due to Rapid Iteration
Quota-related issues with Codex are not unique and will likely continue to occur. The team aims for an infinite supply of tokens and the ability to quickly test multiple ideas, but this compresses the testing period significantly. Many bugs are only discovered during use by users, not before release. This is the reality of cutting-edge AI products: you get the benefits of rapid evolution (such as more intelligent features) but also have to accept that they will always be in a semi-finished state with ongoing bugs and areas needing improvement.
In short, the recent quota crisis at Codex is a consequence of its fast iteration and aggressive collaborative approach. Users can take advantage of the temporary reset options, but they must also tolerate occasional malfunctions—a price of using the most advanced AI technologies.