ChatGPT, Grok, Claude Fail Together: 5 Model Families Hit

For AI engineers and researchers, simultaneous degraded service across ChatGPT, Codex, Claude models and Grok is more than a status-page event—it's a live stress test of the model-serving layer. With Anthropic initially reporting Opus 5, Opus 4.8, Opus 4.6, Mythos/Fable 5 and 5.1 as impacted, and Grok returning 'model overloaded,' the incident raises questions about capacity, shared GPU/inference infrastructure and whether the race to ship new models like OpenAI's Astra is stretching reliability. This briefing digs into the known technical details and what they reveal.

At approximately 11AM ET on September 3, 2026, OpenAI's ChatGPT began returning errors, with the company's status page acknowledging 'elevated errors across ChatGPT and Codex.' Within the same window, Anthropic's Claude and xAI's Grok were already in the midst of their own service disruptions. The near-simultaneous degradation of the three most visible AI chatbots transformed a routine series of status-page incidents into a significant infrastructure story, raising questions about concentration risk, shared dependencies, and the resilience of AI service delivery in an increasingly generative-AI-dependent world.

“

For AI engineers and researchers, simultaneous degraded service across ChatGPT, Codex, Claude models and Grok is more than a status-page event—it's a live stress test of the model-serving layer.

The scope of the ChatGPT incident was broad. OpenAI reported that logins, file uploads, voice mode, search, deep research, image generation, and core conversational use were all affected. The company said it had 'applied a mitigation' and was 'monitoring recovery,' but acknowledged that its tools were still operating with 'degraded performance.' The timing is notable: it comes as OpenAI has been teasing the launch of Astra, a new AI model, and as ChatGPT and Codex have become deeply embedded in developer workflows and consumer tasks. For users and enterprises relying on ChatGPT's function-calling and assistant capabilities, the breadth of the degradation underlined how many features now depend on a single underlying service surface.

Anthropic's Claude outage also had a model-specific footprint. Around 10:30AM ET, Anthropic's status page initially indicated impact across Mythos/Fable 5.1, Mythos/Fable 5, Opus 5, Opus 4.8, and Opus 4.6. As the incident progressed, the company narrowed the issue to Opus 4.8 and Opus 5. Anthropic technical staff member CJ Avilla said an 'infrastructure issue' caused a partial outage across its services. 'We're working on it but don't have an ETA yet,' Avilla wrote on X. That model-level specificity is significant for developers. It suggests that some products and tiers recovered while others remained impacted, and it provides a rare view into how separate model families are served on underlying infrastructure. It also raises a critical operational question: if multiple model versions are served from the same infrastructure, a failure can hit both flagship and legacy tiers simultaneously.

Grok's issues began earliest, at 9:30AM ET, across the app on Android and iOS and on the web. Users attempting to prompt the chatbot on X were met with: 'This model is overloaded right now. Please try again shortly or pick a different model.' The phrasing—'overloaded'—points to capacity or serving-layer pressure rather than a bug in the model itself, and xAI posted a notice saying it was 'working on restoring service as quickly as possible.' Because Grok is tightly integrated into X, the outage also affected the social platform's AI features. Meanwhile, the fact that Grok went down first, followed by Claude about an hour later and ChatGPT about thirty minutes after that, could be interpreted either as a cascading failure pattern or simply as coincidence.

The central unknown is whether the incidents are related. No provider has identified a common cause, and speculation in developer and infrastructure circles ranges from simultaneous surges in demand, to a shared cloud or network dependency, to a coordinated attack. The observable timing is provocative. Three independent companies with different training stacks, model architectures, and serving platforms degraded within 90 minutes of one another. If a shared upstream component—such as a GPU cloud provider, a common CDN, or an authentication service—was involved, it would challenge the assumption that using multiple AI vendors provides true redundancy. For enterprise SaaS products with fallback logic that switches from one model provider to another, simultaneous failure across all three would eliminate the failover path.