OpenAI has once again reset the usage limits for its Codex Proxy and newly launched ChatGPT Work, reflecting the increasing pressure that AI proxies have put on computing infrastructure.
On July 11, Codex Engineering Lead Thibault Sotiao took to X and announced that usage limits for both products would be fully restored for all users within about 30 minutes. He thanked the community for “pushing our systems to the limit,” adding that OpenAI had “never seen traffic increase so quickly.”
Frequent resets indicate a larger challenge beyond momentary service issues. While OpenAI, Anthropic, and Microsoft offer advanced workplace agents, they also face the high processing demand imposed by such systems. These agents are much more expensive to operate because they also perform longer and more complex tasks compared to traditional chatbots.
On Friday, another reset was recorded just one day earlier. On July 10, Sotio stated that OpenAI had already increased the usage limits for Codex and ChatGPT Work while ensuring another reset related to the company’s initial updates.
Hello beautiful people! We’ve reset usage limits across Codex and ChatGPT Work. Another one will come later today. Rejoice.
Now that I have your attention, a quick update on ChatGPT Work, Codex, and all the updates we shared yesterday.
We’ve spent the last 24 hours…
– Thibaut (@thsottiaux) July 10, 2026
The timing is worth noting since ChatGPT Work was only recently launched. According to Fox Business, OpenAI launched ChatGPT Work on July 9, an enterprise-focused agent based on GPT-5.6 that works alongside workplace applications to create reports, spreadsheets, presentations, and other business materials. Releasing such a compute-intensive product at a time when existing infrastructure is already under stress could help explain why users are facing frequent quota resets.
Inside the Codex Bug
The new series of resets follows issues that came to light in the last week of June, when many paying Codex users noticed that their balances were disappearing much faster than expected.
like Cryptopolitan reportedOpenAI attributed the issue to a bug in its fraud prevention system that incorrectly resulted in price caps being applied to certain accounts while using developer credits. Some users claimed their credit consumption rate increased by a factor of 10 to 20 times, while users on the $200 per month Pro plan claimed they saw about $40 in credits disappear within a few hours.
Sotio went on to explain that his team worked over the weekend in what he called a “war room” where they combed through tree trunks and found the problem. OpenAI made three stake resets from June 28-29 before conducting another reset for affected customers.
As it turns out, there was no single cause for the accident. Business Insider noted That Codex was doing much more than expected. Automated code reviews, helper subagents, and retry mechanisms may be triggered multiple times after an error occurs, consuming excessive resources each time. Meanwhile, the usage dashboard showed activity that was never charged for, adding to the chaos.
“All the fixes have now been deployed,” Sotio said after the incident, adding that OpenAI has set up monitoring systems to inform the company if there are any issues in the future. However, recent resets in July suggest that although the glitch has been fixed, capacity issues remain.
The entire sector measures more difficult
OpenAI isn’t the only AI company to have restricted access recently due to increased demand.
Earlier in the year, the Anthropy Index declined Use Claude Limits During high demand for the service, an issue with Claude in March caused outages for developers who began relying on the assistant for programming tasks. The trend now in the industry is that companies are avoiding unlimited access to their services as demand for them has outpaced the available computing power.
The study highlights this matter. In a study conducted by Stanford University’s Digital Economics Lab in April, researchers found that proxy crypto functions require approximately 1,000 times more tokens than traditional crypto conversations, with input tokens largely responsible for this cost. They also found that performing the same task multiple times can result in a significant difference in token usage by up to 30 times despite the fact that increasing the amount of tokens does not necessarily improve the outcome.
These findings highlight why it remains difficult to charge AI agents. Because the cost of computing different tasks can be so variable, providers have very few alternative methods other than usage limits, quotas, or periodic resets to control usage.
But for developers, these measures bring a different set of problems. Many of them now organize their work based on quotas rather than deadlines. OpenAI says it will continue to monitor usage and provide further updates if necessary, but the frequent resets underscore the challenge facing the entire industry: today’s AI agents are becoming more capable faster than companies can build the infrastructure required to support them.
