OpenAI confirmed an active investigation into Codex quota depletion after Pro plan users began exhausting weekly limits in one to two days, a problem that reveals just how fast AI coding agent adoption is compressing demand forecasts.
Something unexpected happened when developers started treating Codex as a real workhorse. They actually used it. Pro plan subscribers paying $200 a month, normally granted quotas designed to last a week, reported burning through their entire allowance in a day or two on tasks that previously stretched comfortably across seven. OpenAI's own status page logged it as an active incident: "Codex Usage Limits Depleting Faster Than Expected." Thibault Sottiaux, the engineering lead on the Codex team, confirmed publicly that abuse and fraud prevention systems had been overflagging ordinary usage, causing certain accounts to drain at abnormal rates. The team applied mitigations and pushed a universal quota reset to all affected users.
The GitHub issue thread told a sharper story than the status page did. One user on the Pro plan reported exhausting nearly their full weekly allowance within about two days of ordinary refactoring work. Others described quota dropping by 15 to 20 percent after just two basic prompts. A separate incident thread documented that since around May 10, some users were getting through only a single long refactoring session before hitting the wall entirely. The official characterization of impact as "limited" sits awkwardly against that volume of reports from paying subscribers.
OpenAI's response escalated through June. Sottiaux announced a one-off "double reset" on June 18, immediately restoring every user's quota and depositing a second banked reset to use later. The company also shipped a rate-limit reset banking feature on June 11, giving Plus and Pro users one free reset at launch. These are patches, not a fix. The underlying pressure they are responding to is more interesting than the quota mechanics.
Here's what the complaints actually show: developers are running Codex hard. Not experimentally, not occasionally, but as a primary tool integrated into real engineering sprints. The Pro plan's 20x weekly quota multiplier, which looked generous when OpenAI priced it, turns out to be insufficient for teams doing serious work. Users running context compaction on large codebases, multi-file refactors, and iterative debugging sessions are burning tokens at a rate the pricing model didn't anticipate. When abuse detection starts flagging legitimate heavy usage as anomalous, that's a capacity model problem masquerading as a fraud problem.
For founders and CTOs building products or internal tooling on top of Codex, the quota reset story is an operational hazard that deserves a line in your risk register. A mid-sprint reset of user caps, applied unilaterally while OpenAI investigates, is exactly the kind of upstream dependency risk that can derail a two-week cycle. OpenAI has been transparent about the incident, but transparency doesn't protect a shipping deadline. Teams relying on Codex for production-speed development should be maintaining API key fallbacks, tracking their weekly burn rates, and not assuming that the quota they start a sprint with is the quota they'll finish it with.
The pricing structure adds another layer of complexity. Codex moved to a token-based billing model in April 2026, with rates varying across input, cached input, and output tokens. A single long refactor session on a large codebase can cost more than most users expect, and context compaction failures, which the GitHub tracker has flagged as a recurring instability, can burn significant tokens without producing any useful output. One thread documented the Plus plan's weekly limit draining in roughly three hours under those conditions. That's a real cost to a small team.
Frankly, the pricing model hasn't caught up with actual usage patterns yet. OpenAI built quotas around what it expected developers to do with an AI coding assistant. What it's discovering is that developers who find a tool useful use it continuously, not intermittently. That gap between modeled demand and actual demand is showing up as an infrastructure and pricing problem right now, but it's also the most concrete signal yet that the enterprise AI coding agent market is developing faster than the incumbents projected.
The competitors will notice. GitHub Copilot, Cursor, and the various API-layer alternatives are all watching OpenAI's capacity struggles with interest. Every Pro subscriber who hits a quota wall mid-sprint is a potential churner, and the churn destination doesn't have to be better, just more predictable. Reliability is a product feature. Right now OpenAI is learning that lesson at scale, in public, in its GitHub issue tracker.
Also read: Asian AI rivals fill the gap left by Anthropic's export ban and match frontier performance for a fraction of the cost • Millennium Management is building its own AI lab and that changes the competitive calculus for quant finance • OpenAI just used AI to build its own chip and that changes the quantum threat to crypto faster than anyone planned
