OpenAI restores Codex and ChatGPT Work usage limits after a traffic increase strained the company’s infrastructure.
On July 11, Codex lead Thibault Sottiaux said on X that limits for both products would be restored for all users within about 30 minutes. He thanked users for “pushing our systems to the absolute limit” and said OpenAI had “never seen traffic increase so quickly.”
The reset followed another adjustment on July 10, when Sottiaux said OpenAI had raised limits for Codex and ChatGPT Work and promised another reset tied to rollout updates.
AI Agents Increase Computing Demand
The repeated changes show how agent-based systems require more processing than chatbots. These tools complete longer, more complex tasks, raising operating costs.
ChatGPT Work launched on July 9. Fox Business described it as a GPT-5.6-based enterprise agent designed to work with workplace applications and produce reports, spreadsheets, presentations, and other business materials.
Its release came while OpenAI was already managing heavy Codex demand. Similar restrictions at Anthropic and Microsoft show that AI providers increasingly rely on quotas and usage caps as computing demand grows.
Codex Users Report Credit Losses
The resets followed a Codex problem discovered during the final week of June. Paying users reported that credits were disappearing much faster than expected.
OpenAI said a fault in its fraud-prevention system incorrectly applied rate limits to accounts while consuming developer credits. Some users said credit use increased by 10 to 20 times. Pro subscribers paying $200 monthly said about $40 in credits disappeared within hours.
Sottiaux said his team worked through the weekend in a “war room,” reviewing logs and identifying the issue. OpenAI reset quotas three times from June 28 to June 29, then completed another reset for affected customers.
Business Insider reported that Codex was doing more work than expected. Automated code reviews, helper subagents, and retries sometimes ran repeatedly after errors. The usage dashboard also displayed activity that was never charged.
Industry Adopts Tighter Usage Controls
“All fixes are now deployed,” Sottiaux said after the incident, adding that OpenAI introduced monitoring systems to flag problems. July’s resets show that the credit bug was addressed, while capacity pressure persisted.
Anthropic also reduced Claude usage caps during heavy demand, and a March service disruption affected developers using Claude for programming tasks.
An April Stanford Digital Economy Lab study found agentic coding tasks use about 1,000 times more tokens than standard coding conversations. Repeating the same task produced token-use differences of up to 30 times without improving results.
These variations leave providers relying on caps, quotas, and resets to manage costs. Developers increasingly organize workloads around limits rather than deadlines. OpenAI said it will keep monitoring usage and provide further updates when needed.
