ChatGPT, Claude and Grok suffered overlapping service incidents on September 3, 2026, but the available provider records do not point to one confirmed common cause. OpenAI attributed the ChatGPT and Codex disruption to a routing error, Anthropic reported elevated errors across Claude services and models, and xAI recorded a separate Grok outage. Reports of Gemini trouble circulated too, but Google did not officially confirm a Gemini incident.
That distinction matters. Three major AI services were unreliable during the same broad period, yet the evidence supports a cluster of provider-specific failures rather than a proven coordinated attack or single cloud outage.
Three confirmed outages, one unconfirmed report
The overlap itself is clear. Anthropic began reporting a partial outage affecting several Claude models at 6:23 a.m. PT. xAI began investigating Grok problems at 6:30 a.m. PT. Around 7:43 a.m. PT, OpenAI said a routing error made ChatGPT and Codex unavailable for some users across platforms.
The incidents did not affect identical products. Anthropic’s main incident covered claude.ai, Claude API, Claude Code and Claude Cowork, with affected model groups including Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8 and Opus 4.6. A separate Claude Sonnet 5 incident was also recorded.
Grok’s incident lasted 3 hours 35 minutes according to xAI’s status record, which later reported that traffic was healthy again at 5:05 p.m. UTC. OpenAI’s initial statement said a solution for the ChatGPT and Codex problem was implemented around 8:17 a.m. PT; its status record later listed the incident as resolved at 4:55 p.m. UTC. Those timestamps describe different stages of the OpenAI response, so they should not be collapsed into one recovery time.
Gemini belongs in a separate category. Users reported problems, but Google did not confirm a Gemini outage in the available public record. It would be inaccurate to describe all four chatbots as officially confirmed down.
What each status page says
The provider-specific record is more informative than the coincidence alone:
| Provider | Affected service | Stated issue | Recovery record |
| OpenAI | ChatGPT and Codex | Routing error beginning around 7:43 a.m. PT | A solution was implemented around 8:17 a.m. PT; the status record later listed final resolution at 4:55 p.m. UTC |
| Anthropic | claude.ai, Claude API, Claude Code and Claude Cowork | Elevated errors affecting multiple model groups | The main affected-model incident was marked resolved |
| xAI | Grok | Service outage under investigation | The incident lasted 3 hours 35 minutes; traffic was reported healthy again at 5:05 p.m. UTC |
Anthropic’s status updates also said the company had identified the cause of elevated errors and was working on a fix. That is a provider-specific explanation for the Claude incident; it does not establish that Claude failed for the same reason as ChatGPT or Grok.
The Claude status history illustrates why outage timing needs context: a dashboard entry shows a partial outage lasting 3 hours 6 minutes on September 3, but that entry should not be treated as the final all-service recovery time for Claude.
Why Grok and Claude looked connected
There is a real relationship between SpaceXAI, Grok and Anthropic. SpaceXAI announced on May 6, 2026, that Anthropic would receive access to Colossus 1 compute to expand Claude capacity. SpaceXAI later attributed the Grok disruption to an outage at its Memphis compute center and apologized to affected compute partners.
That makes a connection between the Grok incident and some Anthropic infrastructure plausible. It does not prove that the Memphis outage caused the Claude incident, and it cannot explain OpenAI’s independently stated routing error. A shared business or compute relationship is not the same thing as a confirmed shared failure path.
The shared-infrastructure theory
AI services can appear independent while relying on overlapping layers beneath the product name. Those layers may include cloud regions, network transit, edge routing, identity systems, DNS, or model-host capacity. If two supposedly separate services depend on the same hidden layer, switching between them may provide less protection than it appears to.
That is a useful reliability model—but it is not a diagnosis of September 3. Cloudflare, Microsoft Azure, Amazon Web Services, CoreWeave and Colossus 1 were discussed as possible shared dependencies, yet the incidents are not attributed to any one of them. The same applies to theories about a coordinated attack or a deliberate test.
A technical analysis of the event presents the “shared basement” idea: different AI brands may sit above common infrastructure, creating what looks like vendor diversity without full architectural independence. The framework is helpful for thinking about correlated failures, while its claims about this specific outage remain speculative.
Another analysis focuses on the Colossus 1 hypothesis and the difficulty of moving large AI workloads between specialized compute clusters. It offers technical context for why frontier-model failover can be difficult, but it does not turn that hypothesis into an established cause of the incident.
What real redundancy would require
For an organization using AI in production, the practical lesson is broader than “keep a second chatbot tab open.” A meaningful failover plan should test whether the backup differs at the layers that can actually fail:
- Cloud platform: Are both services dependent on the same infrastructure provider?
- Region: Would a regional disruption affect both routes?
- Edge and DNS: Do the services share a traffic-management or delivery layer?
- Identity: Can users authenticate if a shared identity system is unavailable?
- Model host: Does the backup require the same scarce compute capacity?
- Application behavior: Can the software switch safely, or would it send uncontrolled traffic to an already busy backup?
The September 3 incidents do not prove that these layers were shared across all three providers. They do show why a list of different AI brands is not, by itself, proof of independent resilience.
The bottom line
ChatGPT, Claude and Grok went through overlapping incidents, but their provider records describe different failure modes: a routing error for ChatGPT and Codex, elevated model and service errors for Claude, and a multi-hour Grok outage that SpaceXAI linked to its Memphis compute center. Gemini was not officially confirmed as part of the outage cluster.
The dramatic single-cause theories remain unproven. The durable takeaway is more useful: AI redundancy depends on separating the infrastructure beneath competing services, not merely choosing a different logo when the first one stops answering.
