Five Major AI Assistants Faltered Together—A Backup Vendor Wasn’t Enough

Five major AI services encountered overlapping trouble on September 3, 2026, but the evidence is not identical for all five. Ars Technica’s same-day account documented disruptions involving ChatGPT, Claude and Grok, elevated errors affecting Codex, and a shorter burst of Gemini problem reports that Google had not acknowledged.

The confirmed incidents were resolved that day. No evidence has established a single cause across the services, yet the overlap exposed a practical weakness in continuity plans that rely only on moving work to another hosted model provider.

Four disruptions were confirmed; Gemini remained a qualified fifth

The strongest provider records cover ChatGPT, Codex, Claude and Grok. OpenAI listed ChatGPT and Codex as affected components, while Anthropic and xAI recognized service problems involving Claude and Grok. That makes these four provider-acknowledged disruptions rather than a collection of user complaints.

Gemini requires more careful wording. External monitoring registered a sharp rise in user problem reports and classified a short API disturbance as a likely outage, but no corresponding Google incident appeared in the evidence reviewed for this story. The headline’s five-service count therefore combines four provider-confirmed incidents with one externally observed disturbance.

That distinction matters when assessing both scale and cause. Monitoring data can show that users encountered failures even when no provider notice appears, but it cannot establish which component failed or whether the disturbance had any connection to the other outages.

The disruption windows overlapped without matching exactly

Claude and Grok were already experiencing problems before the OpenAI incident began. Their recovery times also differed, so “together” describes a consequential period of overlap rather than identical start and end points.

OpenAI’s official September 3 incident record lists elevated errors across ChatGPT and Codex, a mitigation followed by recovery monitoring, and resolution at 4:55 p.m. UTC. It also notes that some Codex remote-control users might need to pair their mobile devices again after the incident.

During part of the overlapping window, a workflow moving from one provider to another could therefore have encountered trouble on both routes. That does not mean every model was unavailable throughout the entire period: depending on timing, product and account configuration, some failover attempts may still have worked.

No common root cause has been demonstrated

The close timing naturally raised the possibility of a shared cloud, network or delivery dependency. WIRED’s account of the September 3 incidents records a routing error affecting ChatGPT and Codex from about 7:43 a.m. Pacific time, an infrastructure issue involving Claude, and a Memphis compute-center outage connected to Grok; it also found no provider evidence establishing one common external cause.

Those descriptions point to different immediate problems, although public incident notices do not necessarily reveal every underlying dependency. The absence of a disclosed common cause does not prove that the systems shared nothing; it means a universal explanation remains unsupported.

Secondary load is another possible mechanism in a multi-provider event. Automated systems may redirect requests when a primary service fails, increasing pressure elsewhere, but the reviewed evidence does not demonstrate that redirected traffic caused or materially extended these disruptions. The defensible conclusion is simultaneity, not shared causation.

Vendor diversity is only one continuity layer

A second model provider remains useful, but this event shows why it is not a complete recovery design. Continuity depends on the entire workflow surrounding the model endpoint, including infrastructure dependencies, retained state and the ability to continue essential work without hosted AI.

• Provider diversity: an alternate model endpoint with credentials, capacity limits and data-handling approval established before an incident.
• Dependency diversity: visibility into the cloud regions, identity services, gateways, network paths and orchestration components used by both routes.
• Cached workflow state: prompts, approved context, intermediate results, job identifiers and audit records retained outside the unavailable assistant.
• Manual fallback: a reduced-service procedure, a responsible decision owner and defined conditions for pausing automation when no model route is dependable.

An alternate model is not an operational substitute if it cannot accept the same approved data, perform required tool calls or resume the state of an interrupted job. Nominal vendor diversity can also hide a common failure point when both routes use the same identity provider, API gateway, orchestration platform or network path.

The relevant planning unit is therefore the business workflow, not the provider list. For each critical process, organizations need an interruption tolerance, a record of which state must remain locally recoverable, and a clear boundary between steps that can continue manually and outputs that must wait for human review after service returns.

Resolution closed the incidents, not the attribution question

By the end of September 3, the provider-recognized disruptions affecting ChatGPT, Codex, Claude and Grok had been closed. Gemini’s shorter disturbance remained supported by external telemetry rather than a Google incident notice, and the reviewed evidence did not connect all five services to one technical failure.

The continuity lesson is narrower than an industry-wide outage but still consequential. Switching vendors can reduce dependence on one model company, yet it cannot guarantee availability during overlapping failures. Resilience also requires independent surrounding infrastructure, recoverable workflow state and a manual operating mode that remains usable when hosted AI routes falter together.

Also read:

• 10 Data Science Beginner Mistakes: Why a Working Model Is Not Enough
• Mozilla Nails It: Thunderbolt Brings “ChatGPT at Home” to the Enterprise — Without Vendor Lock-In
• Claude Stays Ad-Free as ChatGPT Ads Expand Into Five More Markets