At two minutes past midnight UTC on August 20, somebody at OpenAI opened an incident and typed the title straight into the box. It reads, in full: "Chatgpt.com is down - all signups and logins are down as of right now." It stayed open for 52 minutes. The severity field on that incident says minor.
That is not a leak and it is not an accusation. It is a row in a JSON file that OpenAI serves to the public at a stable address, and anyone can pull it in one request. This site did, on the morning of August 22, and then counted what was in it.
What The Feed Actually Contains
The endpoint returns the 25 most recent incidents. The oldest one currently in it opened on July 25, which means the whole of August so far fits inside the window with room to spare. Nothing is being sampled and nothing is being selected. This is every incident OpenAI has declared this month.
Seventeen of the 25 opened in August. All 17 are marked resolved. Here is the shape of them.
Widen the window to all 25 incidents in the feed and the picture does not change. Twenty two minor, three none, and the top two rungs of the severity ladder are empty going back four weeks.
Forty Six Hours, And The Honest Way To Count Them
Adding up incident durations is the wrong way to do this, and it is worth saying so before quoting a number. Several of these overlap. On August 5 three separate incidents were open at once, covering image generation, Custom GPT actions and conversations with files attached. Add their durations together and you get 1,053 minutes out of a day that only has 1,440.
So the count here merges overlapping windows and measures wall clock time instead. How much of August has had at least one OpenAI incident open?
Two thousand seven hundred and seventy nine minutes. Forty six hours and nineteen minutes. Across a 21 day window that runs 30,240 minutes, this is 9.19 percent of the elapsed month.
The naive sum, for the record, is 3,763 minutes. That figure is inflated by roughly a sixth by double counting, which is exactly why it is not the headline here.
It Is Not Spread Out, Which Makes It Worse
Only eight calendar days in August have carried an incident at all: the 4th, 5th, 10th, 11th, 13th, 18th, 20th and 21st. Thirteen days out of 21 have been completely clean.
Look at what that means. Forty six hours of degradation compressed into eight days means that when this platform breaks, it stays broken for a long time. The single longest continuous window opened at 14:04 UTC on August 10 and did not close until 02:48 UTC on August 11. That is 12 hours and 44 minutes, straight through a US working day and out the other side.
Its title is "Increased error rates." Its severity is minor.
August 13 produced a 595 minute window. August 5 produced 507. Three of the four longest windows this month ran past eight hours.
The Word Minor Is Doing A Lot Of Work
Severity grading on a status page exists for one reason. A reader who is not going to read 17 incident write ups needs a single field that tells them whether something is a nuisance or a catastrophe. That field only works if the top of the scale gets used.
Run down what has been graded minor this month. A total signup and login failure at chatgpt.com. Twelve hours and forty four minutes of elevated error rates. Nine hours of conversation failures aimed specifically at free tier users. Codex API authentication breaking. Thinking mode throwing errors for three hours. Unexpected logouts hitting web users for three and a half hours on August 21.
And at the other end of the same scale, graded identically, sits a three minute incident on August 11 in which users could not log in to the ads manager.
A three minute ads console blip and a total login outage receive the same severity label. A scale that cannot separate those two is not a scale. It is a formality.
The Uptime Board, Eighteen Days Later
On August 4 this site pulled the four group uptime figures OpenAI publishes and found ChatGPT finishing last of the four. The numbers that day were FedRAMP at 100 percent, Codex at 99.98, the APIs at 99.93 and ChatGPT at 99.67.
Here they are again, read off the same page on August 22.
Eighteen days, seventeen incidents, and ChatGPT's figure has moved one hundredth of a percentage point. It is still last of the four by a wide margin. The government compliance instance is still perfect. The coding tool is still nearly perfect. The consumer product with the company's name on it is still the worst thing on the company's own board.
Ninety nine point six eight percent sounds fine until you convert it. Against a rolling window it works out to roughly two and a half hours of full unavailability a month, and this month the platform has spent nineteen times that long in some state of not working properly.
Where This Argument Is Weak
Three objections, and all three have real force.
The first is definitional. On a Statuspage style feed, minor has a specific meaning, and it is closer to "partial degradation" than to "everything is fine." Most of these incidents genuinely were partial. Image generation failing while text worked is a real distinction. Thinking mode erroring while standard responses succeeded is a real distinction. Grading those minor is defensible and probably correct.
The second is that publishing any of this is a choice. OpenAI does not have to serve a machine readable incident history to the public at a permanent address, and plenty of companies its size do not. Every number on this page exists because the company decided to expose it. Punishing that is a bad incentive and it should be said out loud.
The third is that 9.19 percent measures the presence of an open incident, not the number of users affected. An incident touching free tier conversations in one region and an incident that takes the whole login path down both count as one open window here. The feed does not publish affected user counts, so nobody outside OpenAI can weight these properly, and this page is not pretending otherwise.
What survives all three objections is the August 20 login incident. Login is broken out as its own component on OpenAI's own status page, and the incident title states that signups and logins were down, not degraded, not slow. That is not partial. If a total failure of the front door is minor, then the word has stopped carrying information, and the two empty rungs above it are decoration.
Three Things That Would Fix This
Publish the severity definitions. Not the vendor's generic ones, OpenAI's own, in writing, so that a reader can check a grade against a rule instead of against a mood.
Weight by users. An incident affecting the login path is not the same size as one affecting the ads manager, and the company has the numbers to say so even if it publishes only a rough band.
Stop self grading, or at least stop grading and reporting in the same breath. Every incident on this page was written, timed, closed and scored by the organisation that caused it. That is the ordinary practice across the industry and it is still a conflict of interest.
Until then, the record stands where the company itself put it. Seventeen incidents. Forty six hours and nineteen minutes. Not one of them major.