AI outages: 105 OpenAI incidents in 90 days, 72 at Anthropic, 6 at xAI and 1 for Gemini

How many outages did ChatGPT, Claude, Gemini, Grok and Perplexity post in 90 days?

Two vendors account for 177 of the 191 incidents we counted, and the gap between them is closing: Anthropic's monthly count fell by about half from August to September 2026 while OpenAI's doubled. We counted every entry on each vendor's official status history on 18 October 2026. One square below is one posted incident.

How we compiled it: we read each vendor's own history on 18 October 2026, OpenAI's 90 day history page and component feed, Anthropic's history and incidents.json, xAI's incidents.json, Perplexity's incident.io feed, and Google's Workspace and Cloud incident feeds. An incident is one posted entry, whatever its label. Scheduled maintenance is left out. Median length is start to resolution as posted, except for OpenAI, where it's the component impact window on the 90 of 105 incidents that carry one. xAI's feed lists nothing before 17 August 2026, so its count may cover a shorter stretch. All times on this page are UTC.

Read the count as a measure of how much each company posts, not only of how often it breaks. OpenAI's page tracks 38 components across ChatGPT, the API, Codex, FedRAMP, an ads platform and the Atlas browser, and it posts notices such as "Delayed support responses" beside real failures. Google's dashboard is the opposite case. It covers Gemini inside Workspace and the Gemini app, and Google Cloud's status feed showed no Vertex AI or Gemini incident at all in the same 90 days.

Is Claude down more than ChatGPT?

No, not on the vendors' own pages: from 20 July to 17 October 2026 Anthropic posted 72 incidents to OpenAI's 105, and its incidents were shorter, a median of 71 minutes against about 1 hour 45 minutes, though OpenAI's figure is its component impact window. Claude's posted incidents were more often model wide, though. Anthropic used its worst label, "Critical", four times: 21 July (twice), 29 July and 16 August. Its 29 July incident title read "Elevated errors across all models".

The trend is the bigger story. Anthropic posted 28 incidents in the last 12 days of July, then 25 in August, 13 in September and 6 from 1 to 17 October. OpenAI's page went the other way. Filed by the date of each entry's latest update, as its history page shows them, it lists 19 in the last 12 days of July, 19 in August, 39 in September and 28 from 1 to 16 October. That filing puts the 29 September outage under 6 October, the day OpenAI posted its report. Many of OpenAI's autumn entries name ChatGPT Work, Codex or Dots rather than plain ChatGPT chat.

User reports tell a different story, and it's worth knowing why. Ookla counted "disruption days" on Downdetector, days when a service drew more than 10 times its usual daily reports. In the first quarter of 2026, Claude had 39 of the 51 such days across four big AI apps, Gemini 7, Copilot 3 and ChatGPT 2, according to Ookla figures reported by Broadband Breakfast on 10 June 2026. A report spike measures how many people noticed. A status page measures what the vendor chose to post. Neither one is uptime.

The closest thing to uptime is a figure only some vendors print. OpenAI's status page showed 99.52% for the ChatGPT group over July to October 2026, with the APIs and Codex at 99.95% each. It says the figures are "reported at an aggregate level across all tiers, models and error types". Perplexity showed 99.95% for its app. Anthropic's page states a 90 day uptime bar but gave no number we could read.

What were the biggest AI outages since July 2026?

The widest AI outage of the period was OpenAI's on 29 September 2026, which hit 30 components across ChatGPT, Codex and the API. OpenAI's incident report gives the cause in one line: "A feature rollout caused routine settings requests to trigger repeated background checks." Those checks used up shared network capacity until healthy services couldn't talk to each other. The report says most of the impact was mitigated within about 30 minutes, by 18:00 UTC, with residual errors until about 23:07. Here are eleven of the biggest incidents the five vendors posted, picked by how much of the main app each one hit, with the cause each vendor gave, if any.

Every cause posted for the incidents in the table was the vendor's own change: a rollout, a routing update and a credential revocation at OpenAI, a rollout at Google, a deploy at Perplexity. OpenAI's Codex write-up says its leak detection system "incorrectly flagged credentials used for communication between internal services supporting Codex as potentially leaked", and a manual revocation then switched them off from about 22:33 to 23:47. Anthropic named no cause for any of its incidents in the table. xAI's page named none for 3 September, though it pointed to a datacenter issue on 22 September and to one GPU cluster on 17 August. The posted length isn't always the felt length, either. Anthropic's 29 July page puts the actual errors at 19:45 to 21:26, about 100 minutes inside a window of 2 hours 47 minutes.

Two long entries are left out of the table because they weren't outages of the chat apps. Anthropic kept "Elevated errors on platform.claude.com" open from 7 to 13 October, six days, and a Cowork on Windows problem ran from 10 to 14 September. OpenAI's longest entries were delayed cost data in its Compliance API, open for up to 70 hours.

Older outages have dropped off every vendor's 90 day history. The one most often cited is Cloudflare's of 18 November 2025, which took ChatGPT and X offline with it, according to Business Today that day. Cloudflare's own report dates the failure from 11:20 UTC, with core traffic largely flowing again by 14:30 and every system normal at 17:06. It traces the cause to a database permissions change that doubled the size of a Bot Management feature file.

What happened in the 3 September 2026 AI outage?

On 3 September 2026 Anthropic, xAI and OpenAI each posted an incident within 77 minutes of each other, and the only cause any of them put on a status page was OpenAI's: a routing error. Here's the morning as their own status pages recorded it, in UTC.

OpenAI's report says its routing system "incorrectly interpreted a configuration update and stopped directing requests to available services." When routes came back, the returning traffic overloaded some services and slowed recovery. Anthropic's page said at 13:41 that it had found the cause. It never named it. xAI posted no cause at all. eWeek reported on 4 September that xAI blamed an outage at its Memphis compute center. eWeek also counted 93 minutes when ChatGPT, Claude and Grok were all affected at once.

Gemini users reported failures that morning too, eWeek wrote, but Google didn't post an incident on its Workspace dashboard or its Cloud status page. Perplexity posted nothing until a billing issue at 20:30. Three vendors failing on one morning looks like one root cause, and the published record doesn't show one. OpenAI's cause was internal, and nobody else posted theirs.

Why do status pages miss outages people can see?

Status pages miss outages because the vendor decides what to post, how to label it and when to start the clock. On 20 August 2026 OpenAI titled an incident "Chatgpt.com is down - all signups and logins are down as of right now", then labelled it "Degraded performance", not an outage. Its write-up says sign-ins failed from about 23:50 UTC on 19 August to 00:19, after a deployment applied "two incompatible configurations" to its login systems. Its 15 October login incident was first posted at 16:29, two hours after the component window it later recorded began at 14:30.

Labels also differ by vendor, so they can't be compared row by row. Anthropic uses minor, major and critical. OpenAI and Perplexity use degraded performance, partial outage and full outage. xAI uses disruption and outage, and Google uses low, medium and high. Google's Gemini app has no status page of its own; its incidents appear on the Workspace dashboard.

So a green page proves less than it looks. If an app fails for you on two networks and the page is green, check a second signal. Our live checkers for ChatGPT and Claude read the vendor feeds and count reports from people who hit the same wall.

What to do when an AI chatbot is down

When an AI chatbot is down, copy your unsent prompt, check the vendor's status page, and switch to another vendor for anything due within the hour. Most posted incidents end soon: xAI's median was 27 minutes and Perplexity's 57. Anthropic's was 71 minutes. If the error names a model, try another model from the same vendor first, since Anthropic's titles often name one model family, such as "Elevated errors for Claude Sonnet 5". Our guide to what to use when ChatGPT is down has the order to try things in.

Whizi won't stop an OpenAI or Anthropic outage. It calls the same vendors, so an incident on OpenAI's API reaches GPT inside Whizi too. What it changes is the switch. On Starter at $15.99 a month, ChatGPT Luna and Claude Haiku sit in one model picker with 280+ models across the plans, so when one vendor posts an incident the other is already signed in. A reply that fails is refunded, not counted against your allowance. How to switch models mid conversation shows the steps.