In the early hours of October 17, Beijing time, people paying for Codex through ChatGPT started reporting that their weekly usage bar had suddenly jumped, wiping out tens of percentage points of allowance without any matching work on their end. A GitHub issue collecting the reports had drawn 37 comments from 26 other users by Sunday afternoon. OpenAI hasn't responded publicly. The only explanation so far comes from support replies that one user posted about their own account, and that user had recently moved to a cheaper plan.
What users are reporting
GitHub user EgorStepkin opened the issue at 03:02 Beijing time on October 17. They wrote that their local Codex logs showed their remaining weekly allowance falling from 77% to 10% in 38 seconds, between 02:10:43 and 02:11:21 Beijing time, while the same logs recorded only about 380,000 tokens in that window, almost all of it cached input. EgorStepkin uses the Codex desktop app on Windows and noted that the figures come from their own logs, not OpenAI's records.
Reports followed from macOS, Linux and Windows, from the desktop app and the CLI. One commenter said their allowance went from 35% left to zero within seconds, and another said their logs showed weekly usage jump from 24% to 97% in 12 seconds at 02:08. Most of the other accounts can only pin the change between two log readings taken minutes or hours apart, and several commenters say outright that they can't show it happened in seconds. The readings they did capture cluster between about 02:08 and 02:12 Beijing time. Some say they were asleep or not using Codex at the time. One commenter pointed out that the affected accounts posted so far all had their next weekly reset scheduled within about 30 seconds of each other, around 14:13 Beijing time on October 22. Several people also say their allowance has run down faster than they expected since then, including after they spent a banked reset to get back to work. All of this comes from users' own logs and observations.
What support told one user
At 12:32 Beijing time on October 17, EgorStepkin posted a reply they said came from OpenAI's AI support agent: "Server-side records show that the weekly allowance used to calculate your meter decreased while the amount already consumed initially stayed the same." According to the reply they posted, support added, "It does not establish that the allowance adjustment was correct." Put simply, the size of their allowance went down while the usage they had already built up stayed put, so the same usage suddenly filled far more of the bar.
EgorStepkin had switched from the $200 Pro plan to the $100 Pro plan on October 6. At 15:30 Beijing time on October 18, they posted a newer reply signed by a support agent named Janis: "The available records confirm that the allowance used by the meter changed on October 16. However, they do not show why the lower allowance was applied ten days after the subscription change, so we cannot confirm whether this resulted from delayed entitlement synchronization or an account configuration correction." October 16 there is a UTC date; in Beijing it was already the 17th. According to the replies they posted, support also told them, "There is no separate published provision describing the exact timing of this type of mid-window recalculation."
That explanation is about their account only. At least three other commenters say they also moved from $200 Pro to $100 Pro in the weeks before the drop, and most commenters who named a plan were on the $100 tier. But EgorStepkin and others in the thread say people who hadn't changed plans were hit too, and the author of a related issue said there had been no plan change on their account that week. Nobody has posted a support explanation for an account like that, so there's no basis yet for saying a delayed plan change accounts for every case. One commenter suspects downgraded accounts had kept the larger $200 allowance until it was corrected, but that's a guess. OpenAI's Pro tiers help article says people who had an active $200 Pro subscription between September 22 and 10 a.m. Pacific on September 29 can keep the previous allowance through October 29 "while their Pro 200 subscription is active."
What OpenAI has said
Nothing in public so far. All 37 comments on the issue come from accounts GitHub shows as having no connection to the repository, and the only other activity there is automated labeling and duplicate notices from a bot. Five related issues filed on October 17, Beijing time, haven't had replies from OpenAI staff either. OpenAI's status page lists no incident about Codex usage between October 16 and 18, and the Codex changelog has no entry on usage limits since its October 15 release. OpenAI's Tibo Sottiaux hasn't mentioned it on X. On the morning of October 17 he posted about a separate change that he said increases the effective usage people get, and his latest global reset finished at 05:12 Beijing time on October 18, but neither post referred to the drop.
If it happened to you
OpenAI's Pro tiers article says support can't reset Codex limits but "can investigate if usage was counted incorrectly." EgorStepkin has asked others who were affected to open their own tickets and mention the GitHub issue. Several commenters spent a banked reset to keep working. OpenAI's help page on banked resets says using a full one refreshes the five-hour and weekly windows and changes your weekly reset date, and that you then don't also get a reset on the original date. At least two of them have asked OpenAI to give back the resets they used if an error is confirmed, and OpenAI hasn't said whether it will. If you keep working after your allowance runs out and have a credit balance, OpenAI says included usage is used first and credits after that, which matches what at least two commenters saw happen to their credits. We'll update this story if OpenAI responds.
