ChatGPT and its coding sibling Codex went down for several hours on September 3, 2026, in one of the largest disruptions OpenAI has logged this year. OpenAI’s status page flagged “elevated errors across ChatGPT and Codex,” and outage trackers lit up almost immediately. By the time the company applied a fix, tens of thousands of people had already filed reports on Downdetector, and the incident had spread across tech coverage worldwide.
The timing mattered. ChatGPT has become a daily tool for hundreds of millions of people, and Codex now sits inside real development pipelines at companies that write code with it every hour of the workday. When both went dark at once, the fallout reached well past casual chatbot users into engineering teams who suddenly lost a tool they had built workflows around.
OpenAI confirms “elevated errors” across ChatGPT and Codex
OpenAI’s own status page carried the clearest account of what was happening in real time. The incident, titled “elevated errors across ChatGPT and Codex,” described degraded performance rather than a full blackout, though for many users the practical effect looked the same: messages failed to send, logins stalled, and coding requests through Codex timed out. At one point during the incident, the OpenAI status page read, “Users across multiple subscription plans may be unable to start or continue tasks in ChatGPT Work. We are continuing work on implementing a mitigation.”
That kind of language, careful and non-committal, is standard for a company managing a live incident in public. It also left a gap that outlets and outage trackers rushed to fill with their own measurements, which is why so much of the public record on this event comes from Downdetector snapshots rather than OpenAI’s own numbers.
Timeline: how the September 3 outage unfolded
Reports on the exact start and end times vary slightly depending on the outlet and time zone used, but the broad shape of the event is consistent. One account put the disruption window between 10:58 AM ET and 2:56 PM ET, a stretch of nearly four hours. Problems began building through the late morning, peaked around midday, and eased into the afternoon as OpenAI worked through a fix.
Users described a familiar set of symptoms: chats that would not load, login screens that spun without resolving, and responses that simply failed partway through. For a product used as often as ChatGPT, even a few hours of instability is enough to generate a large, visible wave of complaints, and that is exactly what happened on September 3.
Downdetector reports climb past 74,000
The scale of the complaint volume is where this outage stood out. One report tracked user-submitted grievances climbing past 30,000 within the first hour alone. Another count, focused on the United States, put ChatGPT-specific Downdetector reports above 35,000. By 8:23 a.m. PT, a separate tally had Downdetector showing more than 74,000 reports tied to the incident, illustrating how fast the numbers grew as the outage stretched on.
Coverage of the event spread quickly to outlets outside the usual US tech press. Moneycontrol reported on users facing problems across both the web and app versions of ChatGPT, while The Times of India and The Daily Star both picked up the story within hours, a sign of how quickly a single vendor’s outage now becomes global news given how many people, in how many countries, depend on the same chatbot.
Which ChatGPT and Codex features went down
The outage was not limited to the main chat window. Multiple reports describe login failures, slow or failed responses, and pages that would not load correctly across both the ChatGPT and Codex product lines. That breadth is part of why the incident drew so much attention: it was not one narrow feature failing quietly in the background, it was the core experience across two flagship products at once.
ChatGPT: logins, chats, and slow responses
For everyday ChatGPT users, the outage showed up as an inability to start or continue conversations, sluggish or stalled responses, and repeated loading errors. OpenAI’s status page acknowledged the scope directly, noting that people on multiple subscription plans could not start or continue tasks in ChatGPT’s work-focused tier while engineers worked on a fix.
Codex: coding workflows interrupted mid-task
Codex users, many of whom rely on the tool inside active coding sessions rather than casual browsing, felt the outage differently. A stalled ChatGPT chat is an inconvenience. A Codex request that fails mid-task in the middle of a build or a pull request review can stall an entire engineering workflow, which is part of why the Codex angle of this story drew as much attention from developers as the consumer-facing ChatGPT angle drew from everyone else.
OpenAI’s mitigation and the eventual resolution
OpenAI said it had applied a mitigation while continuing to monitor recovery, language that matches how the company has handled prior incidents on its status page. The company later confirmed the fix had worked, stating plainly, “This issue has now been resolved.” By the morning of September 4, external trackers and follow-up coverage described ChatGPT and Codex as back to normal operation, with no signs of a recurrence.
What OpenAI has not done, at least publicly, is publish a detailed root-cause writeup explaining exactly what failed. That is not unusual in the first day or two after an incident, but it leaves open questions for the engineering teams who build on top of ChatGPT and Codex and want to know whether the same failure mode could recur.
2026 OpenAI outage timeline
September 3 was not an isolated event. OpenAI logged a smaller, earlier disruption back in April, and the contrast between the two incidents shows how much the scale of complaints has grown as ChatGPT’s user base has expanded over the year.
| Date | Approximate duration | Downdetector reports | Products affected | Outcome |
|---|---|---|---|---|
| April 20, 2026 | Single-evening spike | 1,720 reports around 8:32 PM | ChatGPT | Resolved same evening |
| September 3, 2026 | ~10:58 AM to 2:56 PM ET (about 4 hours) | 30,000+ within the first hour, climbing past 74,000 | ChatGPT and Codex | Mitigation applied, later confirmed resolved |
The April incident, by comparison, was contained and short-lived. The September event dwarfed it in both duration and complaint volume, and it pulled in a second product, Codex, that was not part of the earlier disruption.
Historical context: a rough year for AI platform uptime
OpenAI is not the only AI vendor to have a bad day in 2026. Shattered.io has tracked a run of similar incidents across the industry this year, including a 90-minute outage that hit ChatGPT, Claude, and Grok simultaneously when a shared Azure dependency failed, plus separate cloud incidents at Google Cloud and AWS that took down dozens of downstream services for hours at a time. The pattern points to a wider fragility problem: a handful of cloud providers and model vendors now sit underneath a huge share of daily AI traffic, so a single failure can ripple across products that otherwise look unrelated to end users.
That fragility is compounded by how concentrated usage has become. A few years ago, a ChatGPT outage was an inconvenience for a fairly narrow group of early adopters. In 2026, with ChatGPT integrated into workplace tools, coding pipelines through Codex, and countless third-party apps built on the OpenAI API, an outage of this size touches everything from customer support bots to software release schedules.
How ChatGPT’s reliability stacks up against Claude, Gemini, and Grok
No major AI assistant has a clean uptime record in 2026. All of the leading players, OpenAI included, have had public incidents this year, and several have shared the same root cause because they lean on the same handful of cloud providers underneath.
Shared infrastructure means shared risk
The clearest example is the earlier 2026 incident in which ChatGPT, Claude, and Grok all went offline together for roughly 90 minutes after an Azure-side failure. That event underscored a point that is easy to miss when each vendor issues its own separate outage notice: these products are not as independent from one another as their competing brand names suggest. A cloud-layer failure at one provider can take down services built by entirely different AI labs at the same time.
| Incident | Platforms affected | Reported duration | Scale |
|---|---|---|---|
| ChatGPT and Codex elevated errors | OpenAI (ChatGPT, Codex) | ~4 hours | 30,000 to 74,000+ Downdetector reports |
| Azure-linked multi-platform outage | ChatGPT, Claude, Grok | ~90 minutes | Simultaneous cross-vendor impact |
| Google Cloud regional outage | 33 downstream services | ~2h22m | Broad service disruption |
| GCP us-central1-b outage | 15 downstream services | ~4h08m | Regional infrastructure failure |
| AWS us-east-1 outage | Wide range of dependent services | ~28 hours | Third major us-east-1 failure of the year |
Seen together, the table makes the point better than any single incident does on its own: 2026 has been a year of long, high-profile outages across nearly every major AI and cloud vendor, not a one-off problem specific to OpenAI. What sets the September 3 event apart is the sheer complaint volume it generated in a short window, which suggests ChatGPT’s active user base has grown large enough that even a partial degradation now produces outage numbers that used to be reserved for full platform blackouts.
Market impact: why a chatbot outage moves more than sentiment
A ChatGPT outage in 2026 is not just a consumer story. OpenAI’s products sit inside customer service desks, internal knowledge tools, content pipelines, and, through Codex, active software development. When those tools go down mid-shift, the cost is not abstract. Support queues back up, scheduled content runs get delayed, and engineers lose a tool they had built into their daily workflow.
Enterprise coding pipelines took a direct hit
Codex’s outage is the part of this story that matters most to software teams specifically. Companies that have wired Codex into code review, test generation, or pull request workflows do not have an easy manual fallback for a multi-hour outage. Unlike a chat window, where a person can simply wait or switch tasks, an automated pipeline that expects a response from Codex either stalls or needs a human to step in and do the work by hand until service returns.
That dependency is exactly why enterprise buyers increasingly ask AI vendors pointed questions about redundancy and failover before signing contracts. A four-hour outage that used to be an OpenAI problem is now, for many businesses, also their problem.
Customer support desks felt a similar squeeze. Companies that route a first-response layer through ChatGPT saw queues pile up as the automated tier stopped answering reliably, pushing more tickets to human agents who were not staffed for the sudden overflow. None of that shows up in a Downdetector count, but it is the kind of quiet, operational cost that outages like this one generate away from the headlines.
The single-vendor risk businesses are now reckoning with
The repeated pattern of AI outages through 2026, including this one, is pushing more companies to rethink how much of their stack depends on one vendor. Some teams have started running parallel integrations, wiring both OpenAI’s API and a rival model provider into the same product so a failure on one side does not take down the whole feature. Others have added simple health checks that automatically route requests elsewhere, or fall back to cached responses, when a primary provider starts returning errors.
None of that is free. Running two model providers instead of one adds cost and engineering overhead, and most smaller teams still cannot justify it. But for companies where an outage directly halts revenue-generating work, the calculus has shifted, and 2026’s run of AI and cloud outages is a big part of why.
How to check if ChatGPT or Codex is down right now
Given how often this question comes up during any AI outage, it is worth laying out the fastest ways to confirm a problem rather than guessing. The first stop should always be OpenAI’s own status page, which posts incidents in close to real time. The second is a crowd-sourced outage tracker like Downdetector, which shows a live count of user-submitted reports and can confirm whether a problem is widespread or isolated to one account or network.
Developers who want a quick programmatic check can query OpenAI’s public status API directly instead of loading the page in a browser:
curl -s https://status.openai.com/api/v2/status.json | python3 -c "import sys,json; print(json.load(sys.stdin)['status']['description'])"If the response comes back describing normal operation, the problem is likely local to a single account, browser, or network. Any other response confirms OpenAI is aware of a broader problem and is worth checking the status page directly for details on which components are affected.
What comes next: five predictions for AI platform reliability
- OpenAI will likely publish a more detailed post-incident summary in the days following September 3, following the pattern it has used after past multi-hour outages.
- Expect enterprise customers to push OpenAI and its rivals harder on uptime guarantees and status transparency during contract renewals through the rest of 2026.
- More companies building on Codex or the ChatGPT API will add fallback logic to a second model provider, even if only as an emergency backstop rather than a primary path.
- Outage-tracking sites like Downdetector will keep gaining relevance as a real-time verification layer, since official status pages often lag what users are already experiencing.
- Shared cloud dependencies mean another cross-vendor outage, similar to the earlier incident that hit ChatGPT, Claude, and Grok together, is more likely than not before the end of the year.
None of these are guarantees, but they follow directly from the pattern this year has already shown in shattered.io’s ongoing AI coverage: outages are getting bigger in scale, not smaller, as more of daily work routes through a small number of AI platforms. That trend also sits alongside a separate 2026 story about OpenAI pausing rollout of its Astra model over risk concerns, another sign that reliability and safety scrutiny on OpenAI’s releases has intensified this year.
The bigger picture for AI dependency
What made September 3 notable was not just the outage itself, but how quickly it became a headline outside the tech press, picked up by outlets like MacDailyNews alongside the regional coverage from Moneycontrol, The Times of India, and The Daily Star. A chatbot going down used to be a niche story. Now it is treated as infrastructure news, on par with a cloud region failing, because for a growing share of the working world, ChatGPT and Codex effectively are infrastructure.
That shift in how outages get covered says as much about 2026 as the outage itself does. A tool that answers questions and writes code has become load-bearing for how millions of people work each day, and when it stumbles, the story travels fast.
It also raises a harder question for OpenAI heading into the rest of 2026: growth and reliability are pulling in opposite directions. Every new integration, every enterprise contract, and every developer who wires Codex into a production pipeline adds one more party with a direct stake in uptime. The bigger ChatGPT and Codex get, the less room OpenAI has to treat a four-hour outage as a routine, absorbable event, because the audience watching a status page refresh keeps growing right alongside the product.
Frequently asked questions
What caused the ChatGPT and Codex outage on September 3, 2026?
OpenAI’s status page described the incident only as “elevated errors across ChatGPT and Codex.” The company has not published a detailed public root-cause explanation, so the exact technical trigger remains unconfirmed.
How many users were affected by the outage?
Exact user counts are not public, but Downdetector-style reports climbed from over 30,000 within the first hour to more than 74,000 as the outage continued, according to separate tracking accounts of the incident.
Was Codex affected along with ChatGPT?
Yes. Multiple reports named Codex alongside ChatGPT as affected by the same incident, with developers describing failed or stalled coding requests during the disruption window.
How long did the outage last?
One account puts the disruption window between roughly 10:58 AM and 2:56 PM ET on September 3, a span of close to four hours, though exact start and end times vary slightly across reports.
Has OpenAI confirmed the issue is resolved?
Yes. OpenAI stated, “This issue has now been resolved,” and by the morning of September 4 external trackers showed both ChatGPT and Codex operating normally.
How can I check if ChatGPT is down right now?
Check OpenAI’s status page directly, cross-reference a crowd-sourced tracker like Downdetector, or query OpenAI’s public status API from a terminal for a quick programmatic answer.
Is this the first major ChatGPT outage in 2026?
No. OpenAI logged a smaller outage back in April 2026, and ChatGPT was also part of a separate 90-minute multi-platform outage earlier in the year caused by a shared Azure failure that also hit Claude and Grok.
Could an outage like this happen again?
Given how many AI platforms have had multi-hour incidents in 2026, and how much of that infrastructure is shared across vendors, another large-scale outage before the end of the year is a realistic possibility rather than an edge case.
