Guide · Claude Code errors

What do you do when Claude Code says 529 overloaded?

A 529 means Anthropic's API is at capacity for every user, so it is not your limit and does not count against your quota. Claude Code has already retried by the time you see it. Wait a few minutes, check status.claude.com, or switch model with /model, since capacity is tracked per model. A second Claude account does not help; a different route does.

What does API Error: Repeated 529 Overloaded errors mean?

API Error: Repeated 529 Overloaded errors. The API is at capacity — this is usually temporary. Try again in a moment. If it persists, check https://status.claude.com.

The API answered 529 with overloaded_error: it is temporarily at capacity across all users. Your prompt, key, settings and plan did not cause it.

A 529 is not your usage limit and does not count against your quota. Claude Code retries it on its own, up to 10 times by default, and shows this message only after those retries fail.

The trailing sentence names where to check. On Amazon Bedrock, Google Cloud or Microsoft Foundry it names that provider's status page; behind a gateway, the gateway's host.

Message	What it is	What clears it
`Repeated 529 Overloaded errors`	The API is at capacity for everyone	Time, or another model
`Opus is experiencing high load`	One model is under particularly high load	`/model` to the model it suggests
`You've hit your session limit`	Your plan's own window ran out	The reset time it names; see rate limits

What can you do when Claude Code is overloaded?

Check status.claude.com for a capacity notice, then try again in a few minutes. Retrying at once mostly repeats what Claude Code already did.

To keep working, run /model and pick another model. Anthropic tracks capacity per model, so Sonnet can answer while Opus is overloaded.

To have Claude Code switch for you, set a fallback chain. It tries each model in order when the primary is overloaded, for the current turn only, and keeps at most three:

claude --fallback-model sonnet,haiku

The fallbackModel setting keeps the same chain as an array across sessions. A rate-limit, billing or authentication error never triggers a switch.

How do you keep a CI job running through a 529?

An unattended job cannot wait for you to type again. Set CLAUDE_CODE_RETRY_WATCHDOG=1 and Claude Code retries 429 and 529 capacity errors without a limit instead of stopping. CLAUDE_CODE_RETRY_WATCHDOG_MAX_WAIT_MS caps how long each request waits, so a long outage cannot hold a runner for hours.

Without the watchdog, CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS and CLAUDE_CODE_OVERLOADED_RETRY_MAX_DELAY_MS spread the retries over a longer window.

Good to know: an overload that arrives after Claude finished a block of text or a tool call is not re-run, because that could repeat the same tool calls. Claude Code keeps what was completed and says the response may be incomplete, so check which edits and commands ran before you type "continue".

Why does a second Claude seat not help?

A 529 is about the service, not your account. Every Claude subscription is served by the same capacity, so a colleague's seat or a second Max plan is behind the same wall.

What does help is a different route to a model: a Console API key, Claude on your own Google Cloud project, or a model from another provider. On a team, the developer who hits the 529 is rarely the one who has that set up.

What does Selan do when Anthropic is overloaded?

Selan runs Claude Code and Codex on the subscriptions and API keys your company already pays for. Under Reliability, an owner can turn on Fail over during an Anthropic outage.

With it on, a 529 on a subscription seat is sent again on one of the company's own provider keys instead of another seat, and the developer's session carries on. If the overload arrives after the answer has begun, the resend is spliced into the stream they are already reading. Logs records which account served each request, and whether it failed over.

A second seat is behind the same wall. Selan sends the request on to one of your keys instead.

If the key is overloaded too, Selan answers 503 in its own words, marked as safe to retry, rather than relaying the 529:

Anthropic is overloaded. Selan received this error from Anthropic.

If the key has reached a spend limit instead, the refusal is 402 and names the outage first, so the reader looks at the right key:

selan: Anthropic is overloaded and the fallback key has reached its spend limit — resumes 2026-12-01 00:00 UTC

The model picker also lists models on your other providers, so a developer can move off Claude for the length of the outage. How errors come back is in the CLI reference and the API's error table; every provider is on the models page.

Good to know: a failover can land on a different vendor and region from the seat it left. It only ever goes to accounts you connected, so if residency matters, connect only the ones you accept. The trust page has the detail.

Start for free

Free for up to three seats, and the Free plan needs no card.

The short version

A 529 means Anthropic is at capacity for everyone; it is not your limit and does not count against your quota.
Claude Code has already retried; wait, check status.claude.com, or switch with /model.
For CI, CLAUDE_CODE_RETRY_WATCHDOG=1 keeps retrying, with a cap per request if you set one.
Another seat is behind the same wall; Selan can send the request on your own key instead.

What a 529 means and how Claude Code retries and falls back is as Anthropic's Claude Code error reference and model configuration page (code.claude.com/docs/en/errors and /model-config) stated it on 1 November 2026. It changes: check the date before relying on it. What Selan does is as its gateway, control app and docs stated it on that date.