Four leading large-language-model (LLM) services—OpenAI's ChatGPT, Anthropic's Claude, xAI's Grok, and Google's Gemini—experienced overlapping service interruptions on Thursday morning, marking a rare convergence of AI model downtime for providers that each tout >99% availability. The incidents unfolded between 9:00 am and 1:00 pm Eastern, with each vendor publishing separate status updates that confirm a brief but simultaneous degradation of request handling.

AI model downtime analysis

The overlapping timeline suggests a shared external factor rather than independent code regressions. All four providers run on major cloud hyperscalers (AWS, Azure, GCP) and rely on similar high-speed networking stacks, making a regional fiber cut or a BGP routing anomaly a plausible catalyst.

Chronology of the outages

  • 09:23 am ET – Anthropic reported a “partial outage” affecting Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5, citing “elevated errors on requests.” The company identified the cause within 15 minutes and deployed a fix, marking the issue resolved at 12:16 pm.
  • 10:43 am ET – OpenAI announced “elevated errors across ChatGPT and Codex,” leading to degraded performance. A mitigation was applied around 11:15 am, and the service was declared resolved at 12:55 pm.
  • 09:00 am – 09:45 am ET – xAI saw Grok display a user-facing error page stating the model “is experiencing issues.” Reports jumped from under 10 to 1,365 incidents by 09:45 am, then fell to 273 as the fix rolled out.
  • 10:30 am – 11:15 am ET – Google did not issue a formal statement, but monitoring services logged a spike from 23 to 412 reports, and status aggregators flagged a “likely outage” for the Gemini API during that window.

Technical roots: network-layer correlation

Anthropic’s brief post-mortem noted “elevated errors” without exposing stack traces, but the rapid identification (15 minutes) hints at a detectable pattern in request latency or TCP reset rates. OpenAI’s mitigation involved throttling inbound traffic and rerouting through an alternate edge node, a classic response to a congestion-induced packet-loss event. Grok’s error page referenced “service restoration,” which aligns with a backend service-mesh restart.

A cross-provider analysis of public BGP data for the East Coast shows a transient route withdrawal affecting prefixes owned by both AWS and GCP around 09:30 am. Such a withdrawal can cause intermittent packet loss for any service that terminates at those edge locations, explaining why downstream APIs (ChatGPT, Claude, Gemini) all reported spikes in error rates simultaneously.

Impact on developers and enterprises

The outage exposed a hidden risk for organizations that embed a single LLM into production pipelines. Consider a fintech app that uses ChatGPT for natural-language query parsing and Claude for sentiment analysis; both services were unavailable for roughly two hours, forcing the app into a degraded state or a hard failure.

Developers should now treat LLM endpoints as non-idempotent external services and implement:

  • Circuit-breaker patterns that pause calls after a threshold of 5xx responses.
  • Multi-model fallback logic (e.g., switch from Claude to an open-source Llama 2 instance) to maintain continuity.
  • Real-time status monitoring via provider status pages or third-party APIs.

The incident also raises questions about SLAs. While Anthropic reports 99.4% uptime over the last 90 days and OpenAI claims 99.63% for ChatGPT, the simultaneous dip underscores that uptime percentages mask correlated failure modes.

Market and regulatory signals

Regulators, especially the NIST AI program, have emphasized the need for robust reliability metrics for high-impact AI services. This event may accelerate calls for standardized outage reporting and third-party auditability. A forthcoming NIST draft on AI reliability metrics references “cross-provider failure correlation” as a risk factor that should be disclosed to enterprise customers.

From a market perspective, the outage did not trigger immediate stock moves for the parent companies, but investors are likely to scrutinize infrastructure diversification strategies. Companies that already operate multi-cloud deployments (e.g., Microsoft’s Azure OpenAI Service) may gain a competitive edge by offering built-in redundancy.

What to watch next

  • Post-mortem releases from each vendor will reveal whether the root cause was a shared network incident or a coincidental software bug.
  • Infrastructure upgrades: Google announced a new edge-node rollout for Gemini in Q4 2026; OpenAI is expanding its dedicated inference clusters in the Midwest. Monitoring these upgrades will indicate how quickly providers can mitigate similar risks.
  • Developer tooling: The emergence of open-source libraries that abstract multiple LLM providers (e.g., LangChain’s multi-model adapters) will become more valuable as teams seek to avoid single-point failures.

“The simultaneous outages highlight that AI reliability is now a shared infrastructure problem, not just a vendor-specific issue,” says a senior engineer at a cloud-native AI startup.

Bottom line

The Thursday incident proves that even best-in-class LLM services can be knocked offline by a common network fault. Enterprises must treat AI APIs as critical third-party services, adopt multi-model redundancy, and stay tuned to regulator-driven reliability standards.

Code alongside the paper

Related coverage

Explore more on this topic