There is a particular tell in status page writing, and OpenAI produced a clean example of it this afternoon.

The incident opened at 15:04:53 UTC on August 31 under the title ChatGPT Work seeing elevated errors and latency. The first note said users were experiencing elevated latency and that a mitigation was being worked on. Seventeen minutes later the word latency had become errors, and a sentence had been added: “Plus users are particularly impacted, where Work mode is currently unavailable.”

At 16:02 the scope widened again. “User across multiple subscription plans may be unable to start or continue tasks in ChatGPT Work. We are continuing to work on a mitigation.” Not Plus any more. Multiple plans.

Then at 17:01, nearly two hours in, came the fourth update, and it is the one worth reading twice because of how little it says:

User across multiple subscription plans may be unable to start or continue tasks in ChatGPT Work. We We are continuing work on implementing a mitigation.

Same sentence as the previous update, one duplicated word, no new information. The status is still identified. Not monitoring, which is what a status page says when a fix has been applied. Two hours of a service being unavailable to paying customers, four updates, and the last one is the third one with a typo in it.

What Major Means Here

OpenAI grades its own incidents. August gave it 19 chances to use the word major and it used it once, today.

Severity gradeAugust 2026 incidents
Major1 (the one still running)
Minor16
No grade assigned2

Now look at what got graded minor. On August 20 at 00:02 UTC the status page carried an incident literally titled “Chatgpt.com is down, all signups and logins are down as of right now.” Nobody could log in. Nobody could sign up. Fifty two minutes. Graded minor.

On August 10 an incident called Increased error rates ran for 764 minutes, which is twelve hours and forty four minutes. Graded minor. On August 13, elevated errors in conversations for free users ran 519 minutes. Minor. Image generation on August 5 ran 507 minutes. Minor.

So a total authentication outage is minor, a twelve hour error rate incident is minor, and elevated errors in Work mode is major. The grade is not tracking user impact or duration. It is tracking which customers are affected, and the answer today is the paying enterprise tier.

The Month, Measured

Nineteen incidents, opened between August 4 and today. Total time from open to resolved, with the still open one measured to 17:01 UTC: 4,134 minutes, or 68.9 hours. That is 2.9 days of incident time inside a 27 day window.

Date opened (UTC)MinutesGradeIncident
Aug 10 14:04764minorIncreased error rates
Aug 13 15:19519minorElevated errors in conversations for Free users
Aug 5 14:55507minorElevated errors with image generation
Aug 13 14:03503noneError creating custom RBAC roles for Enterprise users
Aug 5 16:55353minorIssues with Custom GPT actions
Aug 27 18:00254minorErrors in Workspace Agents and ChatGPT Work
Aug 21 19:48207minorUnexpected logouts for some ChatGPT web users
Aug 20 13:56176minorElevated errors for Thinking mode
Aug 4 12:43162minorConversation errors affecting Plus, Pro and Business
Aug 31 15:04116 and countingmajorChatGPT Work elevated errors and latency

The other nine incidents ran between 3 and 127 minutes each.

The Pattern Underneath

Read the August list by product surface rather than by date and it stops looking like bad luck. Work mode and Workspace Agents appear on August 11, August 27 and August 31. Login and session handling appear on August 20 and August 21. Conversation error rates appear on August 4, August 10, August 11, August 13 and August 20.

Those are three separate systems failing repeatedly inside four weeks, and the enterprise surface, the one sold on reliability, is in the list three times including the only major grade of the month.

The thing worth saying plainly: this is not a complaint about downtime. Every large platform has incidents and OpenAI publishes its own, which is more than some do. The complaint is about the grading. A severity label is the number a business uses to decide whether to build a workflow on a service. When a full login outage and a twelve hour error incident both come back as minor, the label has stopped carrying information, and the only signal left is the one nobody wants to rely on: whether your own users are shouting today.

Where It Stands

As of 17:01 UTC on August 31, 2026 the incident is open, the status is still identified, the affected scope has widened once from Plus to multiple subscription plans, and the most recent update repeats the previous one. No estimated time to resolution has been given at any point in the two hours.