OpenAI assembled a warroom over the weekend after a wave of user reports that its Codex coding agent was burning through weekly usage limits in one to two days instead of the usual seven.
The company traced the problem to its own abuse and fraud prevention systems, which were “incorrectly rate limiting certain accounts,” according to an update on OpenAI’s status page. Thibault Sottiaux, the engineering lead for Codex, implemented an across-the-board hard reset of all user caps while the investigation continued, as reported by Business Insider.
“Codex team is in a warroom on a Sunday combing through logs and checking whether there is anything that could lead to increased usage drains for some users,” Sottiaux posted on X. “Taking it very seriously and won’t rest until we get to the bottom of it.”
What Users Reported
Codex usage limits track compute consumption per AI coding task, displayed as a percentage on each user’s dashboard. The limit varies by subscription tier, with more intensive tasks consuming credits faster.
Users across subscription tiers reported the same pattern over the weekend: tasks that previously consumed a week’s allocation were draining it in a day or two.
“I have the $200 plan, and I had to work hard the entire week to burn through the 7 days’ usage,” wrote a software engineer called Adam on X. “The last two days burned the entire week’s usage in one day each, and I had to use reset for the 1st time.”
Another user on the Pro 20x plan reported their weekly quota disappearing in two to three days.
A Pattern Across Providers
The Codex incident follows an outage earlier this month that left users unable to access the tool. Anthropic has experienced similar scaling pressure: Claude usage caps were adjusted during peak hours in March, and a Claude outage the same month highlighted how dependent software engineers have become on AI coding tools, with some lamenting having to write code manually again.
More broadly, AI providers have been moving away from unlimited usage models. Companies have started limiting employee AI access to control costs, and some software engineers have been restructuring their workdays around limit reset schedules.
Reliability as a Competitive Surface
Sottiaux noted the timing: the warroom weekend coincided with OpenAI’s internal “RESET week,” which he described as unrelated to the incident but “funny enough.” The hard reset he implemented overrode users who had stacked up to three banked resets, giving everyone a clean slate.
OpenAI’s status page noted the impact “appears limited” and that monitoring continues. For teams evaluating which agent platform to standardize on, rate-limiting incidents like this one become part of the reliability calculus alongside model capability and pricing.