Codex outage resolved: what to check before changing your login

If Codex stopped working while you were in the middle of a task, OpenAI has confirmed a service outage and says it is now resolved. The incident prompted reports of authentication errors, but changing your entire setup is not a sensible first response to a confirmed provider-side failure.

What has OpenAI actually confirmed?

The official updates move from identifying an internal issue to applying a mitigation and monitoring recovery. OpenAI Developers also confirmed that the service was back. The public record reviewed for this article does not explain the underlying technical failure in detail.

That leaves an important boundary around the story: a login-looking error does not, by itself, establish that your own key was wrong or that your account was compromised. Nor does this incident record establish that every ChatGPT feature failed. We would need separate evidence for either claim.

The API-key workaround has a cost distinction

OpenAI documents two sign-in routes: ChatGPT subscription access and API-key access. Supported local clients can use either, while Codex cloud requires ChatGPT sign-in. API-key use is charged through the OpenAI Platform at standard API rates, rather than deducted from included ChatGPT plan credits.

If you switched during the outage, check the active authentication method before starting a long task. In the CLI, OpenAI documents codex login status for that check. For a work account, follow your organisation’s approved route rather than using a personal key to get around an interruption.

Related: GPT-6 Sol and Luna access and API prices

Resume the task, not just the connection

Our suggested recovery routine is simple: look at the current service status, preserve any useful error message with its time, and try a small request before resubmitting a large job. Remove keys and other private details before sharing logs with support. These are practical precautions, not a claim that we reproduced the incident.

If an assistant was editing files when it stopped, inspect the changes already present before asking it to start over. “Continue from the current files; first explain what is complete and what remains” is a more useful instruction than repeating the original request blindly. A restored connection tells you the service is reachable; it does not review unfinished work for you.

What this says about depending on cloud AI

For deadline-driven work, keep the brief, essential reference files and current output somewhere you can access without the assistant. The useful fallback may be doing one small part yourself rather than moving an entire project into another service in a rush.

If you are considering a tool that can run work on your own machine, read the requirements carefully. “Local” is not a universal promise that every feature works offline. Our Perplexity explainer looks at where local processing ends and cloud assistance begins.

Read next: Perplexity’s local AI for Windows and its hardware requirements