Updated October 11, 2026: this article covers September only. For the longer record, see our dataset of 111 OpenAI incidents from July to October, and for how far the status page’s times can be trusted, our comparison with OpenAI’s own write-ups.
When ChatGPT or Codex breaks, the first question is usually “is this just me?” OpenAI publishes an answer at status.openai.com, and it turns out to be a busier page than most people realise. We read the full incident history for September 2026 and counted 40 separate logged incidents, an average of more than one a day, across ChatGPT, Codex, the API and related services. The four we checked in detail each resolved in under two and a half hours, though we did not check the other 36, and those four turned out not to be typical: our later dataset of 111 incidents (10 July to 8 October 2026) has a median of 114 minutes, with 48% open for more than two hours. We also checked this against what users were reporting on OpenAI’s community forum in the same window, and the two do not always line up, which is worth knowing before you assume a problem is you, or assume it is officially acknowledged.
Key takeaways
- OpenAI’s status history logged 40 separate incidents in September 2026, with at least one resolved on 22 of the 30 days (by UTC date).
- The four incidents we checked in detail lasted between 25 minutes and about 2 hours 12 minutes, and OpenAI published a full root-cause write-up for the largest one.
- A multi-day “Selected model is at capacity” problem that Codex Pro users described in an 87-post community thread does not clearly match any single logged incident of matching length in the same window.
- The status page reports errors in aggregate. An issue affecting only some accounts, as several users reported, may not appear as a public incident at all.
Last reviewed 27 September 2026 and corrected 4 October 2026. The original version counted only the incidents listed through 26 September (37) and quoted clock times exactly as OpenAI’s page displayed them in our browser’s time zone (UTC+5:30), without saying so. This version covers all of September (40 incidents), gives every time in UTC, and fixes two statements: the history page lists each incident’s resolution time, not its start time, and the two 3 September durations we called unreconciled do reconcile (see below). We read OpenAI’s status history, RSS feed, four individual incident pages and one published write-up, plus two threads on OpenAI’s Developer Community forum.
How many incidents were there, really?
Counting from OpenAI’s own status history and its RSS feed (retrieved 4 October 2026), we found 40 separate incident entries resolved in September 2026. The count is 40 whether you group by UTC or by UTC+5:30, and at least one incident was resolved on 22 of the 30 days by UTC date. Every one we saw was marked Resolved by the time we read it. The page shows each incident under the date it was resolved, in the viewer’s own time zone, which is why an incident that ends near midnight can move between months for different readers.
We grouped the 40 by what OpenAI’s own titles say was affected. ChatGPT’s “Work” mode (its agentic browsing and task feature) appeared in 8, and another 8 covered support delays, sign-in, uploads, sharing, advertising login and account creation. Some plan or model explicitly named as affected, such as Plus, Pro, Free or Go, appeared in 7, as did API or backend services such as the Responses API or Agents API, and Codex appeared in 5. Three were described as regional, affecting Europe or the Asia-Pacific region, and two were about image generation.
How long did they actually last?
The history page lists only a resolution time for each incident, shown in the viewer’s time zone, not a start time or a duration. Exact start and end times are on each incident’s own page, and for OpenAI’s most recent 25 incidents also in its public JSON feed, in UTC. We opened four incident pages. Our browser displayed their times in UTC+5:30, so we subtracted 5 hours 30 minutes; for the 14 and 25 September incidents we confirmed the converted times against OpenAI’s UTC feed, and they matched. All times below are UTC.
| Incident | Date and time (UTC) | Length | Notes |
|---|---|---|---|
| Issues with Codex | 25 Sep 2026, 22:58 to 23:54 | 56 minutes | Labelled “critical” on OpenAI’s page. The history page lists it under 26 September in a UTC+5:30 browser |
| Elevated error rates for Codex and ChatGPT Work | 14 Sep 2026, 02:46 to 03:57 | 71 minutes | Checked against the UTC feed |
| Elevated errors across ChatGPT and Codex | 3 Sep 2026, 14:43 to 16:55 | 2 hours 12 minutes | Mitigation applied at 15:17, 34 minutes after the start; see the next section |
| Investigating unexpected usage limit resets | 9 Sep 2026, 17:29 to 17:54 | 25 minutes | Described as affecting some Codex users specifically |
None of the four we checked stayed open longer than 2 hours 12 minutes by the status log’s own timestamps. That matters, because it is still shorter than what some users were describing about their own experience in the same window, which we cover below.
What OpenAI’s write-up says about its biggest logged incident
The 3 September incident is the one OpenAI wrote up in detail (read the write-up, retrieved 2026-09-27), and it gives a shorter duration than the 2 hours 12 minutes between the incident’s first and last status updates. It says the disruption began at approximately 7:43 a.m. Pacific time (14:43 UTC), when “a routing configuration change prevented affected requests from reaching our services”, causing errors including “not found” responses across ChatGPT on web, mobile and desktop, and Codex on the CLI and desktop app, plus some sign-in and account creation errors. OpenAI says the main disruption was mitigated by about 8:20 a.m. Pacific (15:20 UTC), about 37 minutes later, then lists four planned improvements: keeping valid routes available during configuration changes, limiting the reach of routing changes, reducing errors while traffic returns, and detecting routing failures sooner. In an earlier version of this article we said we could not reconcile that 37-minute figure with the 2 hours 12 minutes on the incident page. That was our error. The incident page’s own update timeline reads: investigating at 14:43 UTC, “we have applied the mitigation and are monitoring the recovery” at 15:17 UTC (34 minutes in, matching the write-up), and resolved at 16:55 UTC. The write-up measures time to mitigation, and the status log’s 2 hours 12 minutes measures time to the final resolved mark, so the two figures describe different endpoints of the same incident, with an hour and a half of monitoring in between. The write-up also notes that returning traffic overloaded some services, which prolonged recovery.
The gap between the status page and what Codex Pro users reported
Separately from the status history, we read an 87-post thread on OpenAI’s Developer Community forum, “Selected model is at capacity, please try a different model” (28,800 views, 63 users, 100 likes, retrieved 2026-09-27). Multiple ChatGPT Pro subscribers using Codex, OpenAI’s coding tool, described the error appearing very frequently starting around 10 September 2026, some saying it made their paid Pro accounts “nearly impossible” to use for days. An OpenAI staff member, posting as Prashant_Pardesi, replied on 10 September: “We are looking into this issue and will provide an update soon.” Some users reported the problem continuing into the following days, with one user’s detailed account (in a related thread) describing degraded Codex performance and capacity errors from 10 through at least 15 September, alongside conflicting guesses from users about the cause, including account flagging, IP or proxy use, and account age. None of these guesses were confirmed by OpenAI in what we read.
We could not find a single logged status-page incident covering that full multi-day span at that length. The closest matches by date are the 14 September entry above, at 71 minutes, and the 9 September “usage limit resets” entry, at 25 minutes and explicitly scoped to “some Codex users”. We do not know whether the longer, user-described problem was folded into these shorter official entries, affected only specific accounts in a way that does not get a public incident, or is something else entirely. OpenAI’s own status page includes the caveat that “individual customer availability may vary depending on their subscription tier as well as the specific model and API features in use”, which fits an account-specific problem not showing as a broad incident. We are stating what we found and did not find, not guessing at which explanation is correct.
What this means if you hit an error today
- Check the status page first. status.openai.com/history lists recent incidents with their resolution times. The four we checked in detail each resolved within 2 hours 12 minutes of starting, but across 111 incidents since July the median was 114 minutes and 48% stayed open for more than two hours.
- If nothing is listed, your case may still be real. Our reading shows the page reports aggregate errors, and an account-specific problem may not appear on it at all, so an empty status page does not rule out a genuine problem with your account.
- Search the Developer Community forum at community.openai.com for your exact error message. If others are reporting the same thing with similar timing, it is more likely a broader issue.
- If it is Codex and only your account is affected, one user in the thread we read said switching from Deep Research to Guided Learning worked for them; others reported the same switch made no difference. We found no fix confirmed to work generally, beyond OpenAI’s support channel.
- Report it either way. In the thread we read, OpenAI staff replied after a large number of users had posted about the same error in one place. We don’t know whether the volume of posts caused the reply, but it is where a staff response appeared.
What we could not verify
- We opened only 4 of the 40 incidents to check exact duration. The other 36 are counted from the history page’s titles and the RSS feed’s resolution times only.
- We do not have a reliable way to confirm whether the multi-day Codex capacity reports match, overlap with, or are separate from the shorter logged incidents.
- We read OpenAI’s Developer Community forum, not Reddit or other platforms, so there may be reports elsewhere we did not see.
- This count covers September 2026 by UTC resolution date, as read on 4 October 2026. OpenAI can amend or add entries later.
- Our category groupings are our own reading of OpenAI’s incident titles, not a classification OpenAI publishes itself.
This is the same primary-source method we used for our Gemini Live troubleshooting analysis, and if a Gemini error code is what sent you here, our Gemini error code analysis covers that instead.
How we researched this
On 27 September 2026 we opened status.openai.com/history in a browser and copied the list of incidents. On 4 October 2026 we re-read the history page and OpenAI’s RSS feed, which carries UTC resolution times, to cover all of September, then counted and grouped the 40 incidents by the title OpenAI gave each one. We opened four individual incident pages, chosen for relevance to Codex and to cross-check against the forum thread, and recorded the timestamps shown on each, converting them to UTC. We read the published write-up for the 3 September incident in full. Separately, we read two threads on OpenAI’s Developer Community forum about a Codex capacity error, recording vote and view counts, dates, and any OpenAI staff replies, as displayed on 27 September 2026. Where we describe what happened, it comes from OpenAI’s own page or a forum post; where we note something we could not confirm, we say so.
Frequently asked questions
Check status.openai.com/history, which lists current and recent incidents. In the four incidents we checked in detail from September 2026, each was resolved within about two and a half hours, but that was a small sample. In our later dataset of 111 OpenAI incidents (10 July to 8 October 2026) the median incident stayed open for 114 minutes and 48% for more than two hours, so do not assume a short outage.
The status page reports errors in aggregate across all users, and OpenAI itself notes that individual availability can vary by plan and model. An issue affecting only your account may never appear as a public incident.
We found no OpenAI page that defines the message. In the community thread we read, it was reported mainly by Codex users on Pro plans starting around 10 September 2026, and OpenAI staff said on 10 September they were looking into it. We found no public confirmation of the cause.
By our count, OpenAI logged 40 incidents in September 2026, with at least one resolved on 22 of the 30 days (by UTC date). The four we checked in detail each lasted under two and a quarter hours, and many incidents affected specific features or plans rather than the whole service.
For the same status page counted over a whole month, see ChatGPT Work Mode outage: status said recovered, users didn’t.
For the same count done on OpenAI’s status page, see Claude broken on Windows 4 days: the fix was Microsoft’s.
If something is failing right now, our step-by-step guide shows what to check first: ChatGPT or Claude not working? A five-step check.
For how the two providers compare month to month, see ChatGPT vs Claude: who logged more incidents in September 2026.
If you saw Conversation not found or Error in message stream, see: what OpenAI says and what users report.
For our full count of OpenAI’s incidents that month, see did one outage take down ChatGPT, Claude and Gemini.
Free: the AI Tool Cost Cheat Sheet
What ElevenLabs, Google Vids, Make, n8n and the Claude API really cost, on one page, with every source linked. No sign-up needed.
