Claude internal server error: Anthropic's side, not yours, and the median outage ran 63 minutes

How do you fix a Claude internal server error?

Retry once after a minute, then check status.claude.com before you change anything on your side. Anthropic's Claude Code error reference puts the cause plainly: "A 5xx from the API itself indicates an unexpected failure inside the API. It is not caused by your prompt, settings, or account." So clearing cookies won't fix it. Here's the order:

1. Copy your unsent message. During the 29 September 2026 incident, Anthropic warned that "some messages sent between 14:00 and 14:59 UTC may not have been saved."
2. Stay signed in. The same incident broke sign in, and Anthropic's update said: "If you're signed in, please don't sign out."
3. Open status.claude.com and look for "Elevated errors" on claude.ai, the Claude API or Claude Code. That's Anthropic's name for a wave of 500s.
4. Wait a minute, then retry once. In Claude Code you can type try again, since your original message is still in the conversation.
5. Report it if it lasts with no incident posted. Use Get help in claude.ai, /feedback in Claude Code, or the request ID on the API.

How long does a Claude internal server error last?

Most Claude error incidents end within two hours, and the median one lasts about an hour. On 12 October 2026 we pulled the 50 latest incidents from status.claude.com's public feed, which covers 27 July to 7 October 2026. Twenty six had "errors" in the title. Twenty five of those were resolved, and they ran a median 63 minutes from the posted start to the resolved notice.

The longest was "Elevated errors across many models" on 30 July 2026, at 291 minutes. The shortest, a login failure on 24 August, was posted and resolved within 3 minutes. A stray 500 with no incident behind it is a different case, and Claude Code's own 500 message calls that one "usually temporary".

Don't search the status page for the number. None of the 50 incidents uses the words internal server error or 500, in the title or in any update. Anthropic writes "Elevated errors" instead. An older incident does carry the literal name: Internal server error on Claude.ai, dates from 12 September 2024. Anthropic says claude.ai users got the error for 77 minutes that day.

Will switching to Claude Code or the API get around it?

Rarely. A Claude 500 usually starts in the back end that every Claude app shares, so they fail together. Of the 36 incidents that listed claude.ai, 29 also listed both the Claude API and Claude Code. Here's how often each surface showed up in the same 50 incidents:

The status page's own 90 day uptime, read on 12 October 2026, puts claude.ai and Claude Code at 99.47% and the Claude API at 99.54%. A 0.53% gap sounds small. It's about 11 and a half hours across 90 days.

A different Claude model works more often than a different app. Some incidents name one model, like "Elevated errors for Claude Opus 5.5" on 6 October 2026, which ran 19 minutes. Pick another model in the picker and you're past it. Claude overloaded counts which incidents a model switch fixed, and it covers the 529 capacity error, which is a different problem.

Why does Claude show an internal server error when you log in or pay?

Sign in, checkout and uploads run on the same claude.ai back end as chat, so one incident can break all four. On 29 September 2026 Anthropic reported a second issue "preventing many users from signing in", and the same update listed making purchases and uploading files. That incident ran 126 minutes.

If the status page is green, a login error is more likely on your side. Anthropic's error message guide pairs "There was an error logging you in." with three fixes: turn off any VPN, disable browser extensions, and clear your cache and cookies. For a broken Google sign in, the login help page says to remove Claude at myaccount.google.com/connections and try Continue with Google again.

On a Subscribe button, check before you pay again. Anthropic's billing FAQ sends you to Settings, then Billing, to confirm your recent payment status. A real decline has its own wording, and Claude card declined decodes it.

What should developers do with a Claude 500 api_error?

Retry with exponential backoff and keep the request ID. Anthropic's API error reference, read 12 October 2026, says to contact support "with the request ID" if a 500 persists. Every response carries a request-id header. The Python and TypeScript SDKs expose it as _request_id.

The official SDKs already retry 5xx errors "twice by default" with exponential backoff. Raise max_retries for batch work. Watch streams too: an error can arrive after the API has answered 200, as an error event, so a status check alone misses it.

Claude Code retries up to 10 times before it shows the 500, and CLAUDE_CODE_MAX_RETRIES changes that number. It won't re-run a request that fails after a finished text block or tool call, because that could run the same tool calls twice. You get "The response above may be incomplete" instead, with the finished part kept.

It's a common report. On 12 October 2026 a search of the anthropics/claude-code issues found 882 issues containing "API Error: 500". That's more than three times the 248 that contain "API Error: 529". Only "API Error: 400" turned up more, at 2,177.

When should you report a Claude internal server error?

Report it when the error keeps coming back for more than a few minutes and status.claude.com shows nothing. In claude.ai, click your initials or name in the lower left corner and choose Get help. Anthropic's support page is blunt: "We don't offer phone support." Send the exact error text, a screenshot and the time it started.

In Claude Code, run /feedback. It sends the transcript to Anthropic and offers to open a prefilled GitHub issue. On the API, quote the request ID from the failed response.

What can you use while every Claude app is failing?

Another vendor's model, with your thread pasted in. Keep a short summary of the task in a note (the goal, what you've decided, the last thing that worked) and paste it into GPT or Gemini. What to use when ChatGPT is down lists the other providers and their status pages.

Whizi, which publishes this page, keeps Claude, GPT and Gemini in one chat, and switching models mid thread carries the conversation across (how switching works). A reply that fails is refunded, not charged. Where it loses: Whizi doesn't host Claude itself. Its Claude messages reach Anthropic's models through the OpenRouter gateway, so an incident on a Claude model can fail Whizi's Claude rows too. A Claude model you picked is never swapped for another vendor's behind your back. You get Claude or an error, and the fix is the same switch to GPT or Gemini. Claude Sonnet and Claude Opus start on Pro. Claude Haiku comes with Starter (Claude in Whizi).