Claude Outage Fixed Amid Anthropic's Rapid Claude 5.5 Rollout

An hour offline, and a sign-in problem that lasted longer

On Tuesday, September 29, Anthropic's Claude went down for about an hour in the middle of the U.S. workday. Problems began around 10 a.m. ET, when complaints on Downdetector started to climb. Anthropic acknowledged the issue about 20 minutes later and said nearly every product it sells was affected. Claude for Government was the only exception.1 The status page reported elevated error rates on Claude.ai, including the desktop and mobile apps, and on Claude Code and Claude Cowork. Users saw failed requests, conversations that would not load or send, and unexpected requests to sign in again.2

The outage did not stay limited to the consumer chatbot. As it went on, platform.claude.com and the Claude API also showed degraded performance, which broke developer integrations as well as the chat interface.4 Anthropic's own UTC timeline shows platform.claude.com being added to the list of affected services at 14:30 UTC.7

The recovery came in two stages. Anthropic said it applied a mitigation at 14:36 UTC (10:36 a.m. ET), after which error rates dropped sharply and most existing conversations worked again.2 A second problem remained. Single sign-on and Sign in with Apple were unavailable, and many users could not start new chats, hold voice conversations, open Claude Code or Cowork sessions, upload files or make purchases.210 Anthropic advised users who were still logged in not to sign out.4 The company later said all services had been running normally since 14:59 UTC and that it would keep monitoring.2 One lasting effect: messages sent between 10 a.m. and 10:59 a.m. ET may not have been saved.16

Where the accounts agree, and where they don't

The main facts are consistent across reports: a start time of about 10 a.m. ET, an acknowledgment roughly 20 minutes later, Claude for Government unaffected, and full recovery within about an hour.167 The details differ in a few places.

The size of the outage depends on when Downdetector was checked. One live account counted about 2,900 reports just after 10 a.m. and more than 11,800 by 10:17 a.m.2 Another cited a peak reading of 18,335 reports against a normal baseline of about 10.7 Neither figure is wrong. They are snapshots of a curve that rose fast, and the larger number is the better guide to how many people were hit.

The time of the fix also varies. One early report put the mitigation at 10:40 a.m. ET,8 while later versions of the same coverage and Anthropic's own statement say 10:36 a.m.102 The 10:36 time comes from Anthropic's status update, so it is the one to trust.

Reports also describe the API's condition differently. One described the Console and API as having "degraded performance" while Claude, Code and Cowork had partial outages.10 Another said the trouble had spread across the whole stack, including the API.4 Both can be true. A degraded API still breaks production apps that depend on it, and developers would have felt that even if it was not labeled a full outage.

The ripple effect on rivals was small. ChatGPT and Google Gemini saw outage reports in the hundreds, compared with tens of thousands for Claude.8 That points to a problem inside Anthropic rather than a shared cloud or network failure. It is a different situation from an industry-wide disruption earlier in September that affected several AI providers, including Anthropic.10

No root cause had been published when the incident moved into monitoring. The open question is whether the unsaved messages can be recovered.7

The outage came in the middle of a fast release schedule

The timing matters for anyone following Anthropic's model releases. The outage hit one day after Anthropic launched Claude Sonnet 5.5, the second model in the Claude 5.5 family.1415 Claude Opus 5.5 had arrived a week before that, on September 22.11 Claude Haiku 5.5 followed on October 7, completing the family in just over two weeks.1316

The changes from one model to the next are significant. Anthropic says Opus 5.5 performs about as well as Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.11 Its API price fell to $4 per million input tokens and $20 per million output tokens, down from $5 and $25 for Opus 5.20 Anthropic also says it beats rival frontier models on agentic coding benchmarks at a fraction of the cost per task.18 It was the company's first release since CEO Dario Amodei called for pacing development at the frontier, and it went through external testing by groups including METR before launch.1120 Sonnet 5.5 is described as more than 30% faster than Sonnet 5 and up to 30% cheaper for most work.19 Haiku 5.5 is reported to cost around 75% less to run than Haiku 4.5.16

The pace started well before September. Anthropic's release notes show Opus 4.7 in April, Opus 4.8 in late May, Fable 5 and the restricted Mythos 5 in early June, Sonnet 5 at the end of June, Opus 5 in late July, and Fable 5.1 and Mythos 5.1 on September 1.14 That is roughly one major model every few weeks for half a year.

The same week brought product launches too. Claude for Government, a FedRAMP High–authorized service for federal and state agencies, became generally available around the time Sonnet 5.5 shipped.15 It was also the one product that stayed up on September 29.1 That may simply reflect separate infrastructure built for government compliance. Still, it shows that some parts of Anthropic's setup can be isolated from failures elsewhere.

Reading the pattern: reliability is now part of the product

The coverage does not show that the Sonnet 5.5 launch caused the outage. Anthropic has not linked the two, and the failure looked like a sign-in and session problem rather than a model failure.210 Treating the launch as the cause would be speculation.

The broader pattern is harder to dismiss. Status records from the weeks after the outage show a string of smaller incidents. They include elevated errors on Claude Opus 5.5, separate errors on Mythos 5.1 and Fable 5.1, a delay in crediting accounts on the Claude Platform, and errors loading usage data on platform.claude.com.5 On October 7, the day Haiku 5.5 launched, a spend-limit bug wrongly paused some organizations and caused refused requests across the API, Claude.ai, Claude Code, Claude Security and Cowork.3 The usage-data problem on the developer console lasted for days, although Anthropic said API requests and core functions were not affected.35 One monitoring service measured Claude's 30-day uptime at 99.42%,3 which works out to several hours of disruption a month.

In our view, Anthropic's progress on model quality is now running into a reliability problem. Each new release makes Claude more appealing for heavy, long-running work: million-token context windows, coding agents that run for hours, and subagent setups where Haiku 5.5 supports Opus 5.5 or Sonnet 5.5.1317 Those are the workloads that suffer most when sign-in, sessions or the API fail for an hour. A developer partway through a long Claude Code session gains little from a 40% price cut if the session dies and the work is not saved.

Anthropic does get credit for how it communicated. Reporters noted that its status page acknowledged the problem quickly and was updated at regular intervals, with promised follow-ups every 30 minutes.27 The company has a clear pattern of saying what is broken while it is broken.

What to watch next: whether Anthropic publishes a root-cause report for the September 29 incident, and whether the gaps between launches start to include visible infrastructure work. For now, the company is shipping new models faster than its platform has shown it can run without interruption. Customers deciding whether to depend on Claude 5.5 should weigh uptime as well as benchmark results.