ChatGPT and Codex Fall Over in Broad Thursday Outage
OpenAI's status page lists 15 broken components, from login to Voice mode, as engineers hunt the cause.
Key points
- OpenAI's ChatGPT and Codex services started failing at roughly 10:58 AM Eastern on Thursday, September 3.
- At least 15 ChatGPT components are affected, including login, conversations, file uploads, Voice mode, image generation and the Agent feature.
- OpenAI has acknowledged the outage on its status page and says it is still investigating.
- The company hasn't named a cause, and there's no evidence tying the outage to the upcoming Astra model launch.
ChatGPT is broken for a lot of people right now.
OpenAI's chatbot and its Codex coding assistant started throwing errors at around 10:58 AM Eastern on Thursday, September 3, and hours later engineers were still hunting the cause. Try to open a conversation, sign in, upload a file or run an image generation and you'll likely see a failure.
OpenAI acknowledged the problem on its public status page. "We are investigating the issue for the listed services," the company said. It hasn't disclosed what triggered the outage.
What is actually broken?
Almost everything a normal user touches. OpenAI's status page lists 15 affected ChatGPT components, which is unusually broad for a single incident.
The affected pieces include Conversations, Login, ChatGPT Work (the business tier), Codex inside the ChatGPT desktop app, the Compliance API used by enterprise customers, Search, file uploads, Voice mode, custom GPTs, image generation, Deep Research, the Agent feature that carries out multi-step tasks, ChatGPT Atlas, Sites and Connectors/Apps.
We've tracked Codex across nine stories since first covering it on 1 June 2026, and a full-platform failure of this breadth is new territory.
| Component | Status |
|---|---|
| Conversations, Login, Search | Degraded |
| Voice mode, image generation, file uploads | Degraded |
| Deep Research, Agent, GPTs | Degraded |
| Codex (in ChatGPT Desktop) | Degraded |
| ChatGPT Work, Compliance API, Connectors | Degraded |
Should ordinary users worry about their data?
No evidence points that way. This looks like an availability problem, meaning the service is failing to respond, not a security breach. OpenAI has described it only as "elevated errors" and hasn't mentioned any data exposure.
If you were mid-conversation when the outage hit, unsaved context inside that session may be lost. Either way, don't paste sensitive credentials or customer records into any chatbot.
Is this connected to the Astra launch?
Probably too early to say. OpenAI has confirmed a new flagship model called Astra is coming soon. Big model rollouts often involve serving-infrastructure changes, a common cause of outages. But OpenAI hasn't linked the two, and no public evidence does either. Treat the timing as coincidence until the company publishes a post-incident writeup.
What businesses that rely on it should do now
If your workflow depends on ChatGPT or Codex, assume degraded service for the rest of the working day and plan a manual fallback. Enterprise customers using the Compliance API, which lets regulated firms log and audit ChatGPT usage, should note that audit pipelines may have gaps covering the outage window and verify retention once the service recovers.
Developers piping Codex into their editors should have a local model or cached suggestion path ready. This isn't the first time: a global outage on 20 August 2026 locked users out of ChatGPT, Codex and a dozen API endpoints simultaneously. Betting a production pipeline on a single external AI vendor with no offline mode is a design choice, and days like this are the bill.
As first reported by BleepingComputer, OpenAI is still working the problem. We'll update once the company posts a root cause.
