If you tried to get a response out of ChatGPT this week and got a spinning wheel or a flat error instead, you weren’t imagining it. OpenAI’s chat product, its Codex coding tool, and its API platform all went through a rough stretch between July 23 and July 25, 2026, with error rates spiking, dropping, and spiking again over roughly two days.
I pulled the timeline together from OpenAI’s own status page along with third-party monitors that track it independently, because the status page alone tends to undersell how long these things actually run.
What happened, in order
The trouble started Thursday, July 23, midmorning Eastern time. OpenAI’s status page logged an “Elevated Error Rates” incident affecting ChatGPT, the API, and Codex starting around 11:30 AM ET. The University of Colorado’s IT office, which mirrors OpenAI incidents for its own users, described the cause as an issue with “a downstream infrastructure provider,” which is about as much detail as OpenAI has given publicly so far.
A mitigation went in a couple of hours later, and by early-to-mid afternoon OpenAI was calling it resolved. That didn’t hold. A second wave hit the same evening, starting around 5:56 PM PT and running until roughly 8:41 PM, again labeled as elevated error rates affecting ChatGPT specifically. StatusGator’s independent tracking puts the cumulative bad-service window on July 23 at about eight hours and twenty minutes.
Friday, July 24, brought more of the same rather than a clean recovery. The status history shows a string of separate incidents through the day: elevated API error rates and latency on the gpt-image-2 model just after midnight, elevated errors in Codex code review around 9:35 AM, and another elevated-error-rate incident on the core ChatGPT/API stack starting at 10:06 AM that ran an hour and a half, followed by a much shorter one just after 11:30 AM. Android Authority, which had been following the story since it broke, reported OpenAI calling the whole mess resolved around 3:47 PM ET Friday, roughly 24 hours after the first ticket opened.
Even that wasn’t quite the end. StatusGator’s tracker for the OpenAI API specifically shows warning status extending across essentially all of July 24 and into the morning of July 25 and logged a Codex-specific outage along with entire ChatGPT platform starting around 9:05 AM ET on July 25 that, as of this writing, OpenAI’s own status page has not acknowledged. (Update: OpenAI has updated its status page although no useful information is provided other than components being down)
Live Updates🔴
- 25 July 9:00 UTC – ChatGPT, APIs and Codex Down Worldwide. Logging issues too
- 25 July 9:18 UTC – OpenAI Updates its Status page, Acknowledging the ongoing issue
- 25 July 10:00 UTC – All OpenAI Services Including ChatGPT Back Online
What OpenAI has and hasn’t said
To OpenAI’s credit, the status page updates mostly kept pace with what independent monitors were seeing. StatusGator’s own detection of the July 25 disruption and OpenAI’s official acknowledgment of it landed within the same hour, which at least makes the status page a reasonable first stop when things break.
What’s missing is a real explanation for the July 23 outage. “Downstream infrastructure provider” is the only causal language I’ve found attributed to OpenAI on this incident, and it appears secondhand, through a university IT bulletin rather than OpenAI’s own postmortem. OpenAI has not published an incident report naming the provider, the failure mode, or why the same error signature kept recurring across three separate days instead of getting fixed once. If a written postmortem shows up, I’ll update this piece.
Is this related to the OpenAI/Hugging Face security story?
The same week this outage wave was unfolding, OpenAI separately disclosed what it called an “unprecedented cyber incident” in which models under evaluation in a security testing environment reportedly broke out of that environment and reached into another company’s infrastructure. That story is a genuinely significant security event and deserves its own coverage, which I’ll be doing separately.
But I want to be careful not to blur two different things together. Nothing in OpenAI’s status page history, nor in any of the outage tracking sources I reviewed, ties the elevated-error-rate incidents on ChatGPT and the API to that containment incident. The outage pattern here, recurring elevated errors traced to a “downstream” dependency, reads like ordinary infrastructure issye. I’d treat any social media claims linking the two as unconfirmed until OpenAI or independent reporting says otherwise.
If you’re hitting errors right now
A few practical notes if you’re still seeing failures:
First, check status.openai.com directly rather than relying on secondhand trackers, since OpenAI’s own updates are the fastest-arriving signal even when they’re thin on detail. Second, if you’re building on the API, a simple retry with backoff clears a surprising number of these elevated-error windows, since they tend to be partial rather than total outages. Third, if you’re on Codex specifically, the July 25 morning incident hit that component harder than ChatGPT itself, so don’t assume that because chat works, your coding workflow will too. In fact, currently, almost the entirety of OpenAI is down.
I’ll keep this page updated if OpenAI publishes a postmortem or if the pattern repeats again this week. Three distinct outage windows in 48 hours, on a service this widely embedded in other companies’ workflows, is the kind of thing that deserves a follow-up regardless of how it gets explained.
