At 3:25 PM today, OpenAI opened an incident titled "Elevated ChatGPT conversation errors affecting Plus, Pro, Business, and Edu users." Read the tier list in that sentence again. Plus, Pro, Business, Edu. Every category named is a category that pays. The free tier did not make the headline, which either means the free tier was fine or means nobody thought it was worth naming, and it is not obvious which of those is worse.
That incident is the first of August. July had eleven of them, and eight of those eleven landed in the last eight days of the month.
None of this is leaked, sourced, or disputed. OpenAI publishes it, on purpose, at a public address, and the page is the single most damning document about ChatGPT that anybody has produced this year, precisely because the company wrote it.
The Scoreboard Nobody Reads
The status page groups OpenAI's services into families and prints an availability figure for each across a rolling window running from May to August 2026. Here is the whole board.
Line those four numbers up and the ranking is not subtle. The government-compliance instance is perfect. The developer coding tool is nearly perfect. The API that other companies build products on top of is very good. The consumer chatbot, the product that is synonymous with the company's name and the reason most people have heard of it, is last.
The gap is larger than the decimal points make it look. Availability of 99.98 percent means roughly one part in five thousand goes wrong. Availability of 99.67 percent means roughly one part in three hundred does. Across the same window, the same infrastructure, the same engineers, ChatGPT accumulates something on the order of sixteen times the unavailability that Codex does. On a rolling three-month window, 0.33 percent works out to about seven hours of degraded or dead service. Codex over the same stretch loses under half an hour.
Every user OpenAI has ever advertised to is on the worst-performing tier the company operates. The best-performing one exists to satisfy a federal compliance program.
Eleven Incidents In Eight Days
The uptime percentage is an average, and averages hide clustering. The incident log does not.
Between July 24 and July 31, OpenAI logged eleven separate incidents. July 27 alone produced four: "Elevated errors affecting ChatGPT conversations" at 4:32 PM, elevated latency and interrupted streaming across gpt 5.1 mini and gpt 4.1 mini at 5:30 PM, "Image generation unavailable in ChatGPT" at 7:04 PM, and "Image generation unavailable in ChatGPT" again at 8:24 PM. The same failure, filed twice in eighty minutes, because the first fix did not hold.
July 25 produced two, both titled "Elevated error rates," forty-nine minutes apart. July 28 brought "Elevated Image Generation Error Rate in ChatGPT." July 29 brought a strange one at 2:20 in the morning, "Elevated error rates with the invalid_prompt error code," meaning the system was rejecting user input as malformed when the input was fine. July 30 and July 31 each added another.
Then a three-day quiet stretch, and then today at 3:25 PM.
Read as a list, the pattern is monotonous rather than dramatic. Nothing here is a multi-hour global blackout. These are conversation errors, image generation failures, latency spikes, streaming interruptions, and error codes that fire when they should not. Individually each one is a shrug. Stacked eleven deep across eight days, they describe a product where the median week contains at least one interval during which the thing does not work.
Why The Consumer Tier Is The Broken One
There is a boring engineering explanation and it is probably the right one. ChatGPT has 15 tracked components. Codex has 4 and FedRAMP has 1. A service composed of fifteen moving parts fails more often than a service composed of one, and availability figures compound downward as you add dependencies. ChatGPT carries the image generator, the voice stack, the memory system, file uploads, search, connectors, and the model router, and any of them going down counts against the whole.
That explanation is correct and it is also an indictment, because nobody forced OpenAI to bolt fifteen components onto the product with the largest user base and four onto the one used by professional developers. The FedRAMP instance is at 100 percent because it is deliberately minimal and because a federal customer would escalate. Codex is at 99.98 percent because developers notice and leave. ChatGPT is at 99.67 percent because the people on it mostly refresh the page.
That footnote deserves more attention than it gets. An aggregate across all tiers, models, and error types means the 99.67 percent is a blended figure. If one model family or one subscription tier is meaningfully worse than the rest, the average absorbs it and the page never says so. Today's incident, the one naming Plus, Pro, Business, and Edu specifically, is direct evidence that failures do land unevenly across tiers. The company knows which tiers broke. It publishes one number for all of them.
The Honest Case For OpenAI
Somebody has to say the other half of this out loud, so here it is properly.
Publishing a status page with per-family uptime, a public incident feed, and a footnote admitting the metric is aggregated is more transparency than most software companies offer and vastly more than most AI companies offer. A firm trying to hide reliability problems does not print 99.67 percent next to 100 percent and leave both on the same screen. The disclosure is the reason this article can exist at all, and it should not be treated as a scandal that the company told the truth about itself.
Second, 99.67 percent is not catastrophic by any normal software standard. Plenty of consumer services people rely on daily run worse. The incidents logged in late July were mostly degradations rather than outages, mostly resolved within the hour, and every single one on the page carries a recovery note. There is no unresolved incident sitting open.
Third, the sixteen-times comparison against Codex is arithmetically true and rhetorically loaded. Comparing a fifteen-component consumer surface to a four-component developer tool is not comparing like with like, and any engineer would say so immediately.
The counterargument to all of that is short. OpenAI charges for ChatGPT Plus, Pro, Business, and Edu. Those four names appear in today's incident title. Whatever the architectural excuse, the paying customers are on the tier the company's own instrumentation ranks last, and the tier that runs clean is the one where an inspector general could ask questions.
The Verdict
OpenAI reports 100 percent uptime for its FedRAMP instance, 99.98 for Codex, 99.93 for its APIs and 99.67 for ChatGPT across a May to August 2026 window. Eleven incidents hit in the final eight days of July, four of them on July 27 alone, and an incident opened at 3:25 PM today naming Plus, Pro, Business and Edu users. The company is honest enough to publish the scoreboard and has not explained why the product with the most customers is the one it runs worst.