Omniscio

OmniscioGuides500 Internal Server Error

Claude Code API Error: 500 Internal Server Error

A 500 means your request reached Anthropic's API and the API failed while answering it. Nothing in your prompt, your key or your plan caused it, and the fix is to wait a minute and send the message again.

If you only want the fix, what to do, in order is the section to read. The rest covers what Claude Code already tried, when the number comes from a proxy rather than Anthropic, and how two public reports of it read. Every claim is sourced and dated at the end.

Last checked against Anthropic's documentation on November 16, 2026. By the Omniscio team. Updated .

The message on your screen

API Error: 500 Internal server error. This is a server-side issue, usually temporary — try again in a moment. If it persists, check https://status.claude.com.

The example Anthropic's error reference gives for a 500 on the Anthropic API. The last sentence changes with how you connect, so yours may name a different place to check.

The Claude Code error messages guide covers every other error, from a prompt that is too long to a login that has expired.

What a 500 actually says

Claude Code shows the status code and the API's error message for any 5xx response — the family of server-side failures, 500 and its neighbours such as 502 and 503 — so the text after API Error: is the server's own wording rather than Claude Code's.

Anthropic is direct about what that means: a 5xx from the API itself indicates an unexpected failure inside the API, and it is not caused by your prompt, settings or account.

The API error list puts a 500 under the type api_error: an unexpected error internal to Anthropic's systems, with retrying the request as the response, and contact with support carrying the request ID if it persists.

One part of the message depends on where your requests go. The closing sentence names where to check service health: Amazon Bedrock, Google Cloud's Agent Platform and Microsoft Foundry configurations name that provider's service status, and a custom ANTHROPIC_BASE_URL names your gateway host.

Sources: the server-errors section of Anthropic's Claude Code error reference and the status-code list in Anthropic API errors, read November 16, 2026.

What Claude Code tried before it showed you this

The line is the last thing that happens, not the first. Claude Code retries transient failures up to 10 times with exponential backoff — it waits a little longer between each attempt — and a server error is on that list when it arrives before any of Claude's response has streamed.

There is a narrower retry after thinking begins. A server error that arrives once Claude has finished thinking but before it has started any text or a tool call is retried up to two times; before version 2.1.284, Claude Code ended the turn with the error at that point instead.

Two neighbouring cases are deliberately not retried, and they look different on screen:

  • The failure lands after Claude has completed a block of text or a tool call. Claude Code does not re-run the request, because that could execute the same tool calls twice. It keeps what Claude completed, runs the tool calls that finished, and continues the turn from their results.
  • The failure lands after the response is complete. Nothing needs retrying, so the turn ends normally.

While a retry runs, the spinner counts it down as Retrying in Ns · attempt x/y. For a server error the label reads API error at first and switches to the specific reason from the third attempt, as of version 2.1.198, so a spinner that says only API error early on is normal.

Source: the automatic retries and retry-spinner sections of the error reference, read November 16, 2026.

What to do, in order

Steps 1, 2 and 4 are Anthropic's own for this message; step 3 is its note for the case where every request through a gateway fails this way.

  1. Check the status page. Look at status.claude.com, or the provider status page the message names, for an active incident. A posted incident is the whole answer, and nothing on your machine changes it.
  2. Wait a minute, then send the message again. Your original message is still in the conversation, so for a long prompt typing try again is enough.
  3. If every request fails and you use a gateway, check tool search. A gateway is a server your organisation puts between Claude Code and Anthropic, and you have one if ANTHROPIC_BASE_URL is set. Anthropic's note is to check that the gateway and the service behind it accept the tool search fields Claude Code sends — by default defer_loading tool fields and tool_reference blocks, with the anthropic-beta header that enables them. That check belongs to whoever runs the gateway. Before version 2.1.304, tool search behind a gateway had to be switched on with ENABLE_TOOL_SEARCH.
  4. If it keeps happening with no incident posted, run /feedback. It sends the transcript and a description to Anthropic so the request details travel with the report, and it offers to open a prefilled GitHub issue. Where no Anthropic credentials are configured, it saves a local archive instead.

Two things are worth knowing but are not steps for this error: the request ID in the API's response is what Anthropic's API error list asks you to quote to support, and claude doctor from your shell gives a read-only diagnostic of your installation.

Sources: the 500 entry and the "Report an error" section of the error reference, and the 500 row in Anthropic API errors, read November 16, 2026.

When the 500 is not Anthropic answering

Anthropic notes a case where the number comes from somewhere else entirely. When a proxy, load balancer or gateway answers with an HTML error page of its own, the message shows the status code and that page's title, such as API Error: 502 Bad Gateway.

For a page with no title, it shows the status code and its standard name.

The version matters here. Before version 2.1.281, the status code was dropped when the page had a title, and the page's raw markup was printed when it had none, so an old client can show a wall of HTML instead of a number.

To see which version you are on, run claude --version; claude update moves you to the current one.

If the text after API Error: reads like a web server's page title rather than a sentence from Anthropic, the machine in between is the thing to look at. The connection errors guide covers the family of failures that start there, and the corporate proxy guide walks through a locked-down network.

Source: the 500 entry's gateway paragraph and the Report-an-error section of the error reference; the version flag is in Anthropic's CLI reference. Read November 16, 2026.

500, 529, 429 or a usage limit?

These arrive as different messages and need different responses. The number, or the words, on screen is what tells them apart.

  • 500: an unexpected failure inside the API. Not capacity, not your allowance, and Anthropic's step list offers no model switch.
  • 529 Overloaded: the API is at capacity across all users, after several retries. It is not your usage limit and does not count against your quota, and switching models with /model is one of the documented steps because capacity is tracked per model.
  • 429: the rate limit on your API key, your Amazon Bedrock project or your Google Cloud project. /status shows which credential the request used.
  • A plan's usage limit: a subscription allowance whose message names a reset time, such as You've hit your session limit · resets 3:45pm.

The 529 and 429 share a page: overloaded and 429 errors explained covers the 529 message, the 429 message and fast mode's separate limit in depth. The plan allowance, and what resets when, is on the usage limits page.

Sources: the 500, 529 and 429 entries of the error reference and the status-code list in Anthropic API errors, read November 16, 2026.

What a wave of 500s looks like in the wild

Two issues on Anthropic's public tracker show the shape of a 500 that is not yours. Both were opened in February 2026, and both were closed within hours.

Issue #22838, filed on February 3, 2026, describes 500s on every prompt across all terminals and conversations at once, with the reporter noting normal use seconds before and that restarting the terminal did not fix it. A bot marked it a possible duplicate about four minutes after it was opened, and it was closed the same day.

Issue #28614, filed on February 25, 2026, reports the same thing on a paid plan: a 500 on every prompt, with restarting, new conversations and different prompts already tried. The message in that report carries a request_id, which is the identifier Anthropic's API error list asks for. It was closed about an hour later.

Together they describe a service-side failure: it hits every prompt at once, and the request details the API returns are what a report needs. Neither thread shows a fix a reader can apply, and this page does not claim one.

Sources: both issues, their opening posts, their labels and their close times, read on November 16, 2026.

Four things that do not help

Each of these is a plausible first reaction that Anthropic's steps for a 500 do not include.

  • Reinstalling Claude Code. The step list is the status page, a resend, the gateway tool search check and /feedback; nothing in it touches your installation.
  • Rotating your API key or starting a new account. The failure is decided inside the API, so a new credential changes nothing on the server's side.
  • Switching models. That is a documented step for a 529, where capacity is tracked per model. The 500 list does not offer it.
  • Compacting or shortening the prompt. Request size belongs to the context errors, which are a different family; a 500 is the API failing to answer a request it accepted.

Source: the 500 entry's step list and the automatic-retries section of the error reference, quoted in full above, read November 16, 2026.

Running several sessions at once

When a run fails for good, Omniscio, the desktop app we make for running many Claude Code sessions at once, shows that session as Error in red rather than as something still working, and a message retries it. A failed session collects in the sidebar's Interrupted section instead of nagging from Needs You.

The session statuses page lists all eleven states, and the page on distributing sessions across accounts explains how sessions are spread across your logins. The multi-session guide covers the rest of running several at once.

Where these answers come from

All of the above were read on November 16, 2026. The error reference moves with nearly every Claude Code release and keeps notes such as "Before v2.1.281" and "Before v2.1.304", so it is worth reading again before trusting an older copy of this page.

Common questions

Is a Claude Code 500 error caused by my prompt or my account?

No. Anthropic's error reference says a 5xx from the API itself is an unexpected failure inside the API, and names your prompt, settings and account as not the cause. Nothing you change on your machine reaches that failure.

Should I resend my message after a 500 error?

Yes, after a minute. The original message is still in the conversation, so typing try again is enough for a long prompt. The retries Claude Code makes before it shows the line have finished by then, so a resend is a fresh request rather than a repeat of the last one.

Is a 500 the same as a 529 Overloaded error?

No. A 500 is a failure inside the API, and a 529 is the API being at capacity across all users. Anthropic documents a different response to each: the 529 steps include switching models with /model because capacity is tracked per model, and the 500 steps do not.

Why does the error appear only after Claude Code has been quiet for a while?

Because it retried first. Claude Code retries a transient failure for up to 10 attempts before it shows you an error, so the quiet spell is the retry budget running down rather than the request waiting on nothing.

What is the request ID in a 500 error, and who wants it?

A 500's response can carry an identifier for the failed request. Anthropic's API error list says to contact support with it if the error persists, and Claude Code's own step for a persistent 500 with no posted incident is to run /feedback, which sends the transcript and a description.