Claude Not Working: Elevated Errors Affect Multiple Models

What happened to Claude on September 22?

Anthropic reported an incident involving elevated errors on requests to several Claude models in the early hours of September 22, 2026, in Coordinated Universal Time. The company initially identified three model versions and later confirmed that additional versions in the same families had returned to normal operation.

The disruption was not described as a problem affecting only one person's account or one browser. Anthropic's incident record lists multiple affected services, including Claude's web application, its developer API, Claude Code, and Claude Cowork.

That distinction matters because a user experiencing failed responses during a confirmed service incident may spend time troubleshooting their local setup even though the underlying problem is on the provider's side. Repeatedly signing out, reinstalling an app, or changing browsers cannot directly repair a model-serving issue at Anthropic.

The incident progressed through investigation, identification, partial recovery, monitoring, and resolution. Anthropic's final report states that the impact ended at 02:10 UTC and that the incident was resolved.

Primary source: Anthropic Status — Elevated errors for multiple models.

Is Claude working again?

Yes, Anthropic marked the September 22 incident as resolved. Its status update said that success rates had returned to normal across the affected models and that the company was monitoring the recovery. A subsequent update confirmed the end of the incident.

The exact incident window was 00:50 to 02:10 UTC on September 22, 2026. Anthropic posted its final resolution notice at 02:35 UTC.

An incident being resolved does not guarantee that every user will immediately have a successful request. Someone may have an interrupted session, an expired login, a local network problem, a separate usage limit, or a new service issue. Those possibilities need to be checked individually.

The most useful first step is to open the current Claude Status page. If the page shows another active incident, follow that incident's current updates. If it shows normal operation, use the troubleshooting steps below to determine whether your problem is local or account-specific.

Which Claude models were affected?

Anthropic's first incident notice named Claude Mythos 5.1, Claude Fable 5.1, and Claude Opus 5. At 01:35 UTC, the company reported that success rates for requests to Fable 5 and 5.1 and Mythos 5 and 5.1 had returned to normal. It was still working on errors affecting Opus 5 at that point.

Opus 5 was the model Anthropic continued working on after reporting recovery for the Fable and Mythos versions. At 02:11 UTC, the company stated that success rates had returned to normal across the affected models.

The official incident record does not establish that every Claude model was unavailable or that all requests to the affected versions failed. Elevated errors means requests were failing more frequently than normal; it should not be treated as a claim that every request failed.

Source: Official model-error updates, September 22, 2026.

Did the outage affect Claude Code, Cowork, and the API?

Anthropic's incident record identifies four affected service components: claude.ai, Claude API, Claude Code, and Claude Cowork.

A service appearing in the affected-components list does not mean that every feature within that service was completely unavailable. It identifies the service scope of Anthropic's incident report.

The official record does not list Claude Console as an affected component for this particular incident, although users should check the current status of each service if they encounter a different problem.

Source: Anthropic's official affected-components list.

Official Claude outage timeline for September 22, 2026

Anthropic published five incident updates. The timeline below uses UTC throughout so that the reporting sequence can be compared without mixing local time zones.

The first public investigation update came seven minutes after the start of the impact window identified in the final report. The final resolution notice was posted after the reported impact had already ended.

This is why the time between the first and last status-page posts is not necessarily equal to the period during which users were affected.

Source: Anthropic's complete incident timeline.

How long did Claude's elevated errors last?

According to Anthropic's final incident report, the impact started at 00:50 UTC and ended at 02:10 UTC on September 22, 2026. That is an impact window of approximately 80 minutes.

In Pacific Time, the reported window was 5:50 p.m. to 7:10 p.m. on September 21. The difference in calendar date is a time-zone difference, not a separate outage.

The final status update appeared at 02:35 UTC, approximately 25 minutes after the reported end of the impact.

If you saw a report saying the incident lasted from 00:57 to 02:35 UTC, that period describes the span between the first investigation notice and the final resolution notice. Anthropic's final report gives the narrower 00:50 to 02:10 UTC window as the actual period of impact.

What caused the elevated Claude errors?

Anthropic said at 01:17 UTC that it had identified the cause of the elevated errors and was working on a fix. However, its public incident record does not provide a detailed technical explanation of that cause.

It would therefore be premature to attribute this specific incident to an infrastructure provider, a new Claude model release, a routing change, a security event, or unusually high user traffic without a supporting statement from Anthropic.

The confirmed facts are that request error rates were elevated on several models, Anthropic identified the cause internally, recovery occurred in stages, and the company eventually marked the incident resolved.

Primary source: Anthropic's incident report and cause-identification update.

What might users see when Claude experiences elevated errors?

A service-side model error can appear differently depending on whether a person is using the Claude website, a desktop application, Claude Code, or a developer application connected through the API.

For a chat user, a request might fail before an answer appears, an answer might stop unexpectedly, or the application might ask the user to try again. A developer may instead see a failed request, server-error response, or timeout.

These are possible symptoms of model-request failures in general. Anthropic's September 22 incident report does not publish an exhaustive list of the exact error messages every affected user received.

The important first distinction is whether the error coincides with a provider incident. If it does, local troubleshooting may have little effect until the affected service recovers.

What to do if Claude is still not working after the outage

Anthropic has marked the September 22 incident resolved, so a new error should be investigated using current information rather than assuming it is part of the same disruption.

The following steps help distinguish a new service incident from a browser, account, connection, or usage-related problem.

1. Check Anthropic's live status page. Visit status.claude.com and look for an active incident affecting the specific Claude service or model you are using.
2. Confirm which error you are receiving. An elevated server error, login error, usage-limit message, and capacity warning do not necessarily have the same cause or solution.
3. Save important work before retrying. If a long prompt or partial answer remains visible, copy it somewhere safe before refreshing the page or starting another session.
4. Try a simple, low-risk request. Once the status page shows normal operation, test whether Claude can answer a short message before resuming a complex workflow.
5. Check your connection and browser if the problem appears local. If Claude works on another device or network, investigate the browser session, extensions, VPN, or connectivity.
6. Check your account and usage state. If the message explicitly says you reached a usage limit, follow the limit information shown in Claude rather than treating it as a service outage.
7. Contact support for a persistent unexplained problem. Include the error text, relevant time, affected service, and the troubleshooting steps already attempted.

If a new system-wide incident is active, waiting for Anthropic's restoration updates is generally more useful than repeatedly uninstalling and reinstalling Claude.

Official troubleshooting: Anthropic — Troubleshoot Claude error messages.

Should you switch to a different Claude model when one fails?

Switching models may help when an incident affects only a particular model and another available model is operating normally. However, this is not a guaranteed workaround for an incident that affects multiple models.

During the September 22 disruption, Anthropic initially identified errors affecting Mythos 5.1, Fable 5.1, and Opus 5. Its later update reported that Fable 5 and 5.1 and Mythos 5 and 5.1 had recovered while Opus 5 still had remaining errors.

A user who had access to an unaffected or already recovered model might have been able to continue certain tasks on that model. That would not mean the output, capabilities, context, or behavior were identical to those of the original model.

Before switching, check whether the alternative model is available under your account and suitable for the work you are doing. Avoid switching in the middle of a sensitive or irreversible workflow merely to force a response.

What developers and Claude Code users should do during model errors

Developer applications and coding workflows can require additional care because a failed model response may occur in the middle of a multi-step process. An automatic retry is not always harmless, especially when the earlier request may already have triggered an external action.

For Claude API applications

Check the HTTP status, response body, request ID, and timestamp. Compare the failure against Anthropic's current API incident information before changing application configuration.

For requests that are safe to retry, use bounded retries with backoff instead of sending an unlimited stream of immediate requests. Avoid retrying an action that could create duplicate records, send the same message twice, or repeat a payment unless the workflow has appropriate safeguards.

If your application supports an alternative model, validate that the alternative is available and compatible with the task before routing work to it.

For Claude Code

Inspect the state of the repository and any partially completed changes before resuming. If a request failed while Claude Code was editing files or invoking tools, do not assume that nothing happened merely because the final model response was interrupted.

Check the affected service status, review local changes, and then resume the task from the actual project state. Keep important work under version control so interrupted operations are easier to inspect and recover.

For Claude Cowork

Review any files or external services that the interrupted task may have changed. If an action may have completed before the model response failed, verify the result before requesting the action again.

Anthropic's September 22 incident affected multiple model-backed services, but its public report does not provide a separate technical postmortem for each Claude Code or Cowork workflow.

An outage is not the same as a usage limit or capacity warning

Claude can display several types of errors that feel similar to the user but represent different conditions. Anthropic's help documentation distinguishes service incidents from normal capacity management and account usage limits.

Anthropic's help center states that capacity constraints may not appear on the public status page because they represent normal load management rather than a technical outage. It also describes the specific messages users may see when they reach a plan's usage limit.

If Claude displays a reset time for a usage limit, waiting for an outage to be marked resolved will not necessarily remove that separate account restriction.

Source: Anthropic's guide to usage limits, capacity warnings, login errors, and service incidents.

What happens to usage and billing when a Claude request fails?

An interrupted request can raise questions about usage limits, API charges, and whether an unfinished task should be retried. The September 22 incident report confirms the model errors and recovery timeline, but it does not announce a general refund, automatic credit adjustment, or universal billing rule for every affected request.

For an API application, review the request result and the usage or billing information available in your Anthropic account. For an individual Claude plan, check the usage information shown in the product if a failed attempt appears to have affected your remaining allowance.

If you notice an unexpected charge or believe a failed request was counted incorrectly, preserve the timestamp, model name, error details, and relevant request ID where available. Contact Anthropic support through the account's official support channel for an account-specific review.

Do not assume that every failed request is charged, that no failed request can incur usage, or that the incident automatically entitles every user to a refund. Those conclusions require the applicable billing rules and the details of the individual request.

Support information: Anthropic — How to get support.

How to check whether Claude is experiencing another outage

The September 22 event is now part of Claude's incident history. If you encounter new errors, use the live service status rather than a historical article to determine whether a fresh incident is in progress.

1. Open Claude Status.
2. Check the current status of the service you are using, such as Claude's website, API, Claude Code, or Cowork.
3. Open any active incident and read the latest update. Look for its affected models, start time, and whether Anthropic is investigating, working on a fix, monitoring, or has resolved the issue.
4. Compare the incident with your own error and time of occurrence. An unrelated issue affecting another service or model may not explain your particular failure.
5. If no active incident explains the problem, follow the product's normal troubleshooting steps or contact support.

Anthropic's status page also offers incident notifications, allowing users who depend on Claude for work to receive updates when an incident is created, updated, or resolved.

Official sources: Live Claude service status and the resolved September 22 incident report.