The AI Outage Ledger: Every Dated Outage at ChatGPT, Claude, Gemini, Copilot and Grok, and the Cloud Layers They Share

ChatGPT, Claude and Grok, three direct competitors, failed in the same morning window on September 3, 2026, and the shared-layer record that explains it appears on none of their own status pages, because each vendor logs only itself. At the peak, Downdetector counted more than 37,000 ChatGPT reports, 1,365 for Grok, 1,324 for Claude, and only about 500 for Gemini, the one assistant that does not sit on the affected stack 1. This ledger joins the timelines: every dated outage and degradation we could verify across ChatGPT, Claude, Gemini, Copilot and Grok, compiled from the vendors' own UTC-stamped histories, published uptime figures and dated outage reporting, with each multi-provider event traced to the layer underneath. Figures and entries are current as of October 15, 2026.

Two root causes for September 3, both on the record

Outage tracker shattered.io, citing reporting from Tech Times, places the common trigger in a regional failure inside Microsoft Azure's East US infrastructure and lists Grok's primary cloud host as "Microsoft Azure-linked infrastructure"; it reports OpenAI detected elevated errors across 15 ChatGPT components and 4 Codex components at 10:58 a.m. UTC, with broad recovery by 12:42 p.m. ET 1.

The operator of Grok tells a different story. "We are sorry for the issues you may have experienced with Grok following an outage at our Memphis compute center this morning," SpaceXAI wrote on September 3. "We'd also like to apologize to our impacted compute partners" 2. Engadget reports Grok was down three and a half hours from 6:30 a.m. PT, that Claude started failing around the same time and was resolved by 9:16 a.m. PT, that OpenAI resolved by 9:55 a.m. PT without naming a cause, and that Anthropic signed a deal earlier in 2026 to lease compute from SpaceXAI 2.

The disagreement is the point. If Memphis is the truth, Claude's exposure runs through the same facility as Grok. If Azure East US is the truth, three rivals shared one cloud region. OpenAI's own status messaging never named Azure, according to the coverage reviewed 1. An enterprise mapping its redundancy needs to know which map is real, and the vendors' own pages do not say.

The shared layers under five assistants

September and October show how often the failing layer belongs to someone other than the assistant on the error screen:

• September 3: ChatGPT, Claude and Grok degraded together and Microsoft's Copilot was flagged for stability issues, attributed to Azure East US by tracking coverage and to SpaceXAI's Memphis compute center by SpaceXAI itself 1 2
• September 15: one Claude model-error event appears in two vendors' timelines, with Anthropic opening an investigation at 09:48 UTC and GitHub opening its own at 09:47 UTC and naming degraded Claude Fable 5.1 inside Copilot a minute later, "caused by a problem with our upstream model provider" 3 4
• September 17: Copilot's GPT-5.6 and GPT-6 model surfaces failed between 20:26 and 21:17 UTC on "an issue with an upstream model provider," the same evening OpenAI's own history logs elevated error rates across API models 4 5
• September 14: Claude Cowork on Windows broke because of a Microsoft Windows update shipped September 8, and was repaired by Microsoft's own patch, KB5129195 3
• September 16: two Google-layer failures in one day, with a Google-confirmed Google Play issue breaking Claude subscriptions while Copilot's Gemini 3.8 Flash degraded in a separate incident 3 4

October kept the cadence, layer by layer:

• October 1: GitHub attributed Actions job-start delays to "throttling within an upstream Azure dependency," the cloud at the center of the September 3 dispute, this time named in a vendor's own post 4
• October 5: the Claude model pair from September 15, Mythos 5.1 and Fable 5.1, logged elevated errors again, with a vendor-stated impact window of 12:40 to 13:10 UTC 3; the same day GitHub's Actions degradation, which failed 47 percent of workflow runs on hosted runners at its peak, also brought intermittent errors in repository lists, licensing and billing pages, and some Copilot features 4
• October 7: two GitHub write-operation outages totaling about 19 minutes listed Copilot among the affected services, with 5.6 percent of web and API requests failing in the first window and more than 12,000 organizations affected, per the vendor's own summary 4, while Anthropic's spend-limit system incorrectly paused some organizations, refusing requests on the Claude API, Claude.ai, Claude Code, Claude Security and Claude Cowork 3
• October 8: Gemini failed for some users for roughly an hour with more than 4,000 Downdetector reports, while Google's status page kept showing all systems operational, per dated reporting; next-day complaints centered on Gemini 3.1 Pro 6 7
• October 13: OpenAI's GPT-6.1 Sol ran degraded across ChatGPT Work and four Codex surfaces from 08:10 to 16:36 UTC, a published window of 8 hours 26 minutes 8 5
• October 15: OpenAI logged and resolved a "Widespread feature degradation" the same day 5, while a Compliance API cost-data delay identified October 14 was still open at refresh time, with the vendor stating API and ChatGPT usage are not affected 9

What October did not produce, on the readable vendor pages and in the dated reporting we found, is a repeat of the joint failure: the ledger's October rows below each name a single assistant, where September 3 named four.

Claude's headline year predates the cluster: an April 15 outage that peaked above 30,000 reports and registered 40 minutes of major outage plus 73 minutes of partial outage, and a June 16 outage that peaked at 2,103 reports with Sonnet and Opus models at roughly 10 percent error rates 10 11. Anthropic's April 23 engineering post-mortem traced a six-week quality drift to changes shipped between March 4 and April 16 10. Earlier entries on other stacks: an April 20 routing-layer partial outage at OpenAI that peaked at 8,700 UK reports, and a February 27 Gemini API error spike on Vertex AI's global endpoint 10.

What 99.53 percent uptime buys

As of October 15, 2026, OpenAI's status page puts ChatGPT's trailing-90-day uptime at 99.53 percent and its APIs at 99.96 percent, for the July-to-October window 9. This ledger's September edition recorded 99.63 and 99.93 for the same two surfaces: over the month the consumer side lost a tenth of a point while the API side gained three hundredths, and the component roster shifted underneath, from 15 to 16 ChatGPT components and 12 to 13 API components, with the standalone Sora row no longer listed. The arithmetic the page does not do: 0.47 percent of 90 days is about ten hours of expected unavailability for ChatGPT, while the APIs' 99.96 percent works out to roughly fifty minutes. The history shows where it goes, with OpenAI posting 30 incident entries in the first 19 days of September 2026 and 39 across the full month 5. In the ten days from September 10 to 19, the first window where all three published histories were directly comparable, OpenAI logged 20 incident entries, Anthropic 5 and GitHub 5, and Anthropic's six September incidents include the September 22 model-error event, resolved after a vendor-stated impact window of 00:50 to 02:10 UTC 5 3 4. The 20-versus-5 gap is entry density, not outage density: OpenAI's history logs billing, support and account-side entries alongside availability incidents, from delayed support responses to overbilling for OpenAI-hosted containers, while the Anthropic and GitHub histories carry availability incidents only 5. October ran the comparison again and the pattern held: in the first fifteen days the three histories logged 28 OpenAI entries against 6 Anthropic and 11 GitHub incidents, with two of GitHub's naming Copilot features or services explicitly 5 3 4.

The number an enterprise actually needs

A team running rival assistants as redundancy has, on this record, stacked correlated bets. Copilot's flagship models are OpenAI's models, and its Claude option rides Anthropic's models, as GitHub's own upstream-provider attributions show: the September 17 incident names the OpenAI models Copilot serves, and the September 15 post names Anthropic's Claude Fable 5.1 4. Claude's compute includes capacity leased from SpaceXAI 2. The one assistant structurally outside the September 3 blast radius was Gemini on Google's own cloud 1. And recovery is not instant anywhere: across five September incidents spanning both stacks, Claude's September 10 API-latency and September 11 model-error events and GitHub's September 13, 17 and 20 incidents, the median gap from the vendor's first post to its resolved post is 50 minutes, computed from the vendors' own UTC stamps 3 4. That is the real fallback budget.

October stretched the same measure. Across the fourteen October incidents on the two vendors' pages where both the first post and the resolved post are stamped, the median gap grew from September's 50 minutes to 89, and both ends of the distribution moved: Claude's October 5 and 6 model-error events closed in 16 and 19 minutes, while the Claude usage-data incident that opened October 7 at 13:25 UTC did not close until October 13 at 18:00 UTC, six days later, with core functionality unaffected throughout, and GitHub's October 12 incident was mitigated at its core 20 minutes after the first post yet took until 20:16 UTC, just under eight hours, to clear its notification backlog 3 4. OpenAI's October entries publish resolved stamps but not investigating stamps in our reads, so the October measure covers the two vendors whose timelines expose both. The budget is not the median; it is the spread, and October widened it.

Method, briefly: incident counts come from the vendors' status histories; resolve times from their UTC-stamped first and resolved posts; report volumes are Downdetector figures as published by the cited outlets. Grok's own status page refuses automated readers with an HTTP 403, so every Grok row rides on dated reporting of that page rather than our own fetch of it, and is labeled as such; the April and February entries, the June 16 outage and the September 3 Anthropic window likewise ride on dated reporting of the vendors' pages, as their brackets show. The October 8 Gemini entry rides on dated reporting for the same reason: Google's consumer status surface would not load for our fetcher this cycle after two attempts, Google Cloud's health dashboard showed no broad severe incidents as of October 15 12, and the cited reporting itself notes Google's page showed all systems operational during the event. No dated October outage reporting for Grok was found. Figures and entries are current as of October 15, 2026.