Blog

On 14 September 2026, OpenAI opened a status incident for ChatGPT Work Mode whose very first public update already said a mitigation had been applied and service had recovered. A forum thread started about an hour later shows ChatGPT Plus users still reporting broken tools, failed file creation and image-path errors. OpenAI then posted a second round of mitigations, and users began reporting recovery within about half an hour. This is a case study in the gap between what a status page says at a given moment and what users are seeing, with every time converted to UTC so the sequence can be checked.

Key takeaways

1. OpenAI’s Work Mode incident (opened 14 Sep, 15:58 UTC) began with an update saying a mitigation had been applied and service had recovered. There was no earlier “investigating” update.
2. Between 17:08 and 19:34 UTC, 11 users in a community forum thread reported continuing failures, including missing file-write ability, image-path errors and failed uploads.
3. OpenAI posted “additional mitigations” at 19:42 UTC. Three users reported tools working again from 20:11 UTC, the first 29 minutes later, and the incident was marked resolved at 03:11 UTC on 15 September, about seven hours after the first recovery report.
4. OpenAI Support replied in the thread on 16 September (13:41 UTC), saying the issue should be resolved and asking anyone still affected to open an individual support case.

Last reviewed 4 October 2026 and corrected the same day. The first version of this article quoted clock times as OpenAI’s status page displayed them in our browser’s time zone (UTC+5:30) without saying so, said we had read the whole thread when the page had hidden its last six posts, and concluded that users stayed broken for about a day after the incident was marked resolved and that no OpenAI staff replied. None of that was right. This version uses UTC times from OpenAI’s status feed and the forum’s own data, and covers all 26 posts, including three from OpenAI staff accounts.

What OpenAI’s status page recorded

OpenAI’s page shows two incidents touching Work Mode in this window. The first, “Elevated error rates for Codex and ChatGPT Work“, ran from 14 September 02:46 to 03:57 UTC, 71 minutes, and appears to have ended several hours before the forum thread’s author says their problem began. We cannot tell whether the two are related. The second, “Elevated errors affecting Work Mode in ChatGPT“, is the one the forum thread refers to (data from OpenAI’s UTC incident feed and the incident page, retrieved 2026-10-04):

Two details stand out. The incident opens at the “monitoring” stage, so the public record contains no moment at which OpenAI said it was still investigating. And the thread’s author says their problem began at about 4 to 6 a.m. Pacific time, which is 11:00 to 13:00 UTC, so OpenAI’s incident was opened between roughly 3 and 5 hours after that. The status page’s own caveat applies here: “Individual customer availability may vary depending on their subscription tier as well as the specific model and API features in use.”

What users reported in the forum

The thread (Work Mode tools remain unavailable after Sep 14 mitigation, started 14 September 17:08 UTC, 26 posts, 4,023 views, retrieved 2026-10-04) was opened by a ChatGPT Plus user, 1 hour 10 minutes after OpenAI’s first update. Their report listed a “Some tools are temporarily unavailable” warning on every turn in Work Mode, a Markdown file request that produced no downloadable file, and, on all five PNG uploads they tried, an error reaching the underlying tool: “Codex could not read the local image at /workspace/scratch/…/upload/<file>.png: No such file or directory (os error 2).” They also reported that video analysis that had worked before no longer did, and that the iOS app hid the warning while the same failures continued.

Eleven users in total described continuing failures between 17:08 and 19:34 UTC. One said their session reported no “local file write capability”. Another said Excel editing was inactive. A third said image generation with a reference image failed with “Could not load the image path”. One could not upload PPTX and ZIP files together. One user, after clearing their cache and moving work to a new session, quoted the model’s reply: “I can access the required source files, but this Work session currently lacks the local file-processing runtime needed to extract ZIPs, render exact charts, and build/validate new ZIP archives.” Another wrote: “Putting in a huge amount of effort into a Project only to have Product Failure is unacceptable for PAID SERVICE!”

The sequence, side by side

What this does and does not show

The record supports a narrow claim. For at least about three and a half hours after OpenAI’s first update said service had recovered, users in the thread were still reporting broken Work Mode tools, and OpenAI then posted a second mitigation, which fits the first one not having fully worked, although OpenAI does not say so. Recovery reports followed the second mitigation closely, with the first 29 minutes after it. The final “resolved” mark came about seven hours after the first recovery reports, so in this case the formal resolution was later than what at least three users saw.

The forum user who noticed that the page showed both “resolved” and “monitoring” was probably looking at two separate incidents, one already resolved and one still being monitored, although we cannot tell which they saw. Two more users reported problems on the evening of 15 September, after the incident was resolved, but their symptoms differ, so we do not treat that as the same fault returning. This extends what we found in our count of OpenAI’s September incidents: the page records OpenAI’s own monitoring at the time, and a user’s experience can lead or lag it.

What we could not verify

• We did not experience the fault or test Work Mode ourselves, and we cannot confirm any user’s account of their own symptoms.
• The thread’s author gives their own estimate of when the problem began. We have no independent measure of when it started or ended for any user.
• OpenAI does not explain what the first or second mitigation changed, or why its incident opened at the “monitoring” stage.
• We cannot tell whether the earlier 14 September incident (02:46 to 03:57 UTC) is related to this one.
• The thread has 26 posts from 14 distinct non-staff users plus staff accounts. The forum itself reports 15 participants, a difference we could not explain.
• We read one thread. Other users may have had a different experience that was never posted.

How we researched this

On 4 October 2026 we read OpenAI’s status incident for Work Mode through its public incident feed, which gives update times and text in UTC, and read the earlier 14 September incident’s times from the same feed. We then read the forum thread through its public JSON, which lists all 26 posts with UTC timestamps and marks staff accounts, and compared each post’s time with the status updates. Quotes are copied as posted. Where we say something is a possibility or a fit with the evidence, that is our reading, not a statement from OpenAI or the posters.

For how the same provider compares with another over a month, see our comparison of ChatGPT and Claude incident records, and for a different outage story checked against primary sources, see our look at the 3 September explanation.

Frequently asked questions