ChatGPT, Claude and Grok Went Down at the Same Time on September 3

Several major AI services experienced overlapping problems on September 3, 2026: OpenAI’s ChatGPT and Codex, Anthropic’s Claude, and xAI’s Grok. The failures were conspicuous because users of competing platforms reported trouble within the same morning window.

The most important conclusion is also the most cautious one: these were concurrent disruptions, not a proven single outage. OpenAI identified a routing error in its own service, while the public reporting and provider notices available for the other services did not establish a shared technical root cause. 4041

What the timeline shows

OpenAI said a routing error began at approximately 7:43 a.m. PT on September 3, making ChatGPT and Codex unavailable for some users across platforms. It said a solution had been implemented by about 8:17 a.m. PT and was being monitored. 40

OpenAI’s public status incident was created at 14:58 UTC and described “elevated errors across ChatGPT and Codex.” 17 A later account of the providers’ resolved notices placed OpenAI’s recovery at 16:55 UTC. Measured from the reported 7:43 a.m. PT start time—14:43 UTC—the incident ran for about two hours and 12 minutes, though the impact was partial rather than uniform for every user. 4041

Anthropic and xAI had issues in the same broader period. Reporting based on their status notices said Anthropic’s model-error incident was resolved at 16:16 UTC, while xAI later reported Grok traffic healthy at 17:07 UTC. 41

What users and status pages reported

Users reported failed logins, chat errors, unavailable responses and trouble accessing web and app experiences. Downdetector data cited during the incident showed especially heavy ChatGPT reporting: more than 35,000 U.S. reports, alongside more than 1,400 for Claude and more than 1,200 for Grok. In India, the same report cited more than 2,800 ChatGPT reports, about 170 Claude reports and more than 70 Grok reports. These are user reports, not a measurement of every affected account, but they indicate the disruption was widely noticed. 42

OpenAI’s status page confirmed elevated errors across ChatGPT and Codex, and its status history later marked the incident resolved. 1720 Contemporary reporting described broad ChatGPT impacts across web, mobile and desktop, rather than a failure isolated to one interface. 3950

For Claude, reporting cited elevated errors affecting several models, including Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8 and Opus 4.6. Later updates indicated that many models had returned to baseline while Opus 4.8 and Opus 5 remained affected for longer. 364749

xAI acknowledged that Grok was experiencing issues, and subsequent reporting said its traffic was healthy again after the incident. 4941

Was Gemini the unaffected alternative?

Early accounts often framed Google Gemini as the major rival that appeared to remain available while ChatGPT, Claude and Grok were disrupted. That characterization should be treated carefully. Other coverage reported Gemini-related Downdetector complaints, but the supplied evidence does not show a corresponding confirmed Google incident. 4448

In practical terms, a service can appear usable to one group of users while another group sees elevated errors. Outage trackers are valuable signals, but they are not substitutes for a provider’s incident report or application-level monitoring.

Why the incident mattered beyond chat

For many teams, the impact was not limited to asking a chatbot a question. When coding assistants or agent workflows are unavailable, development tasks can stall. When an API returns errors, product features and internal automations that depend on it can queue up, fail or require manual handling. Research workflows are also exposed when the tools used for search, synthesis or long-running tasks become unavailable.

OpenAI’s own history illustrates the breadth of services that can be affected in AI-platform incidents: separate recent notices have involved Codex, search, agents, Deep Research, connectors, file uploads, login and GitHub-dependent workflows. 19242529

That does not make any one provider uniquely unreliable. It shows why a dependency on one provider, one model or one agent runtime can become an operational single point of failure.

The Astra/GPT-6 speculation was not evidence

The timing fueled speculation because reports connected the outage to rumors of a forthcoming OpenAI announcement referred to as “Astra” and potentially associated with a future GPT-6-era model. But the available evidence does not support a causal connection.

OpenAI’s reported explanation for its own disruption was a routing error. No provider publicly tied the simultaneous problems to Astra, GPT-6 or a model launch. The overlap made for an attention-grabbing narrative; it did not establish causation. 403746

What prior incidents say about reliability risk

The September 3 event was not the first outage in this market. xAI’s status history includes a Grok-2 timeout on January 15, 2025 that lasted 3 hours and 10 minutes, as well as a structured-output outage for Grok-2 that lasted 3 days and 21 hours. 23 OpenAI’s status history likewise documents distinct incidents involving Codex tasks, ChatGPT functionality and API-related workflows. 212732

These records should not be used to rank providers by reliability: the available evidence does not provide comparable definitions of downtime, exposure or affected users. They do support a narrower conclusion: outages are a normal planning risk for teams that build critical processes on hosted AI services.

A practical continuity checklist

Organizations that depend on AI for production work can reduce the blast radius of provider incidents by preparing for failure before it happens:

• Use provider-neutral interfaces where practical, so an application is not tightly coupled to one vendor’s request format or workflow.
• Define fallbacks for important tasks, whether that means a second provider, a smaller local model, a reduced-feature workflow or a manual process.
• Queue and retry safely for non-urgent API jobs, with limits that prevent repeated failures from amplifying an outage.
• Separate critical workflows from convenience features. A chat assistant may be optional; a workflow that blocks code review, customer support or a product action may not be.
• Test failover and manual runbooks. A backup that has never been exercised is not a dependable continuity plan.

The September 3 disruptions were unusual because several well-known AI services appeared to fail at once. Their larger lesson is ordinary but important: resilience depends less on predicting the next outage than on ensuring one provider incident cannot stop essential work. 41