ChatGPT restored after four-hour outage

ChatGPT is back up, but the interesting part is what actually broke. This was not just a case of the chatbot spitting out bad answers or slowing down a little. OpenAI’s status page shows a Sunday evening incident that hit file uploads inside ChatGPT and task creation in Codex Cloud, then took a few hours to fully clear. (status.openai.com) By late evening Pacific time, OpenAI marked the incident resolved. ### What happened? OpenAI logged an incident on Sunday, May 10, 2026, saying users could see elevated errors when uploading files in ChatGPT and when creating tasks in Codex Cloud. Those are two very specific product actions — not vague “degraded performance” in the abstract. (status.openai.com) The company later said it had applied a mitigation and was monitoring recovery, then marked everything fully recovered at 10:35 PM Pacific. ### How long was ChatGPT affected? The timestamps matter here. The incident window on the status page runs from 6:23 PM to 10:35 PM Pacific on May 10. That is a little over four hours from first logged impact to full resolution. (status.openai.com) OpenAI also posted a monitoring update at 8:23 PM, which suggests the fix or workaround landed well before the final all-clear. ### What exactly was broken? The clearest symptom was file uploads in ChatGPT. If you rely on ChatGPT to read PDFs, spreadsheets, screenshots, or other attachments, that path was the one failing. At the same time, Codex Cloud task creation was also affected, which points to a problem that crossed product surfaces. (status.openai.com) In plain English — this looked bigger than a single chat window glitch. ### Why does the Codex part matter? Because it hints that the outage sat somewhere in shared infrastructure or a dependency layer, not only in the model stack people usually think about. ChatGPT and Codex are different user experiences, but they can still depend on common services for uploads, job handling, orchestration, or storage. (status.openai.com) OpenAI did not publish a root cause in the incident note, so that last step is still an inference — but the cross-product blast radius is hard to ignore. ### Was all of ChatGPT down? Not necessarily. The status page labels this one as “degraded performance,” not a full outage, and it lists one affected ChatGPT component rather than the whole service. (status.openai.com) That usually means some users or workflows were failing while others kept working. OpenAI also notes on its main status page that availability can vary by subscription tier, model, and feature set, so two users can experience the same incident differently. ### Why do file-upload failures feel worse than they sound? Because uploads are a gateway feature. If that step fails, a lot of higher-level use cases fail with it — document analysis, spreadsheet help, image interpretation, project work, and anything built around attached context. (status.openai.com) It is a little like a hotel whose elevators stop working. The rooms still exist, but a big chunk of the building suddenly becomes impractical. The model may be fine, but the product no longer feels fine. ### Is this unusual for OpenAI? No — and that is the broader takeaway. OpenAI’s status history shows a steady drumbeat of narrower incidents around file uploads, transcription, model-specific errors, and API hiccups, even while the overall status page now says systems are fully operational. (status.openai.com) That is normal for a large live service, but it also shows where reliability work has to happen — in the seams between features, not just in the core model. ### So what’s the bottom line? The outage is over. But the lesson is bigger than one bad Sunday night. ChatGPT is no longer just a chatbot — it is a bundle of upload, retrieval, tasking, and orchestration features layered around models. (status.openai.com) When one of those layers breaks, users experience the whole thing as “ChatGPT is down,” even if the model itself is still there. (status.openai.com)