Claude overloaded error 529: switch models first, then vendors

What fixes Claude overloaded in the next 60 seconds?

Switch the chat to a different Claude model and retry once. If the status page names several models, move the thread to another vendor. The first step works because, in the words of Anthropic's Claude Code error reference, capacity "is tracked per model". So a full Opus says nothing about Haiku. In order:

1. Copy the unsent message out of the box before you touch refresh.
2. Read the wording against the table below. A capacity or overloaded message is not your usage limit; a message with a reset time is.
3. Switch models. In claude.ai, click the model name next to the send button; Anthropic says the change applies "starting with Claude's next response", so the thread stays. In Claude Code, run /model.
4. Open status.claude.com. An incident naming one model confirms step 3. One naming "multiple models" means every Claude model is affected, so take the thread to GPT or Gemini.
5. Retry once after a minute, not twenty times. Claude Code has already retried up to 10 times before it shows you the error.

Is Claude overloaded right now, or is it just you?

Check status.claude.com (status.anthropic.com redirects there), which posts incidents separately for claude.ai, the Claude API, Claude Code, Claude Cowork and the Console. A green page doesn't clear Claude, though. Anthropic's error message guide, updated 18 May 2026, says the capacity message is load management: "Capacity issues will not appear on our status page because they represent normal load management rather than technical problems." A capacity message on a green page is the short kind. A posted incident is the long kind, and we measured how long.

On 3 October 2026 we read every incident on status.claude.com's history for August and September 2026. Of 38 incidents, 24 were model errors or degraded model performance. The rest were logins, billing, connectors and single apps. Twelve titles named specific models, such as "Elevated errors for Claude Sonnet 5". They ran a median 35 minutes from the posted start to the resolved notice, never longer than 95. The other 12 said "multiple models", "all models" or every Claude service, and ran a median 112 minutes. The longest, on 5 August, lasted 429 minutes. When the incident names models, another Claude model is the fix. When it doesn't, waiting costs about two hours. The "2 to 5 minutes" other guides quote describes only the capacity blips that never reach the status page.

Timing matters too. Thirteen of the 24 began on a weekday between 9 a.m. and 6 p.m. Eastern, a window holding 27% of the week's hours, so US working hours carried about twice their share.

Which Claude model should you switch to when one is overloaded?

Claude Haiku 4.5 is the safest switch. It was named in 1 of the 12 incidents that named specific models in August and September 2026, against 7 for Claude Sonnet 5. Count from status.claude.com incident titles, read 3 October 2026. One incident can name several models, so the column adds up to more than 12.

The pattern fits Anthropic's own advice. Claude Code's message for a busy Opus tells you to "use /model to switch to Sonnet". When Sonnet is the one in trouble, Opus or Haiku is the move. Pick by the job: Haiku for a rewrite or a quick answer, Opus for the hard part, and come back to your usual model on the next message.

What does the Claude 529 overloaded_error mean?

A 529 overloaded_error is the Claude API saying it has no capacity for your request right now. The request itself is fine. Anthropic's API error reference, read 3 October 2026, defines it as "The API is temporarily overloaded" and warns that 529 errors "can occur when the API experiences high traffic across all users." Three neighbours look alike and need different fixes:

One trap in streaming code: an error can arrive after the API has already answered 200, as an error event inside the stream. If you only check the HTTP status, you'll miss an overload that lands partway through the answer.

Does an overloaded error count against your Claude usage limit?

No. Claude Code's error reference says it in one line: "A 529 is not your usage limit and doesn't count against your quota." Your usage limit is a token budget that resets five hours after the session's first message. Its wall names a reset time, and Claude usage limit reached covers that one. An overload names no time, because nothing about your account changed.

Paying more buys priority, not immunity. Claude Max lists "Priority access at high traffic times" on Anthropic's pricing page, read 3 October 2026, and it's the only plan line Anthropic sells against capacity. Every incident in the count above was posted against claude.ai, the API or Claude Code as a whole, with no plan carved out. Claude Pro vs Claude Max weighs what else the bigger plan buys.

How do you make Claude Code and the API fall back on their own?

Give Claude Code a fallback chain and give your API client a model list, so an overload becomes a switch instead of a failure. Claude Code takes claude --fallback-model sonnet,haiku for one session, or "fallbackModel": ["claude-sonnet-5", "claude-haiku-4-5"] in settings to keep it. Per Anthropic's model configuration docs, it tries the entries in order when the primary model is overloaded or unavailable. It caps a chain at three models and switches for the current turn only, so the next message tries your main model again. Nothing confirms the chain at startup. The first notice you'll see is the switch itself.

Two more settings matter for long jobs. CLAUDE_CODE_MAX_RETRIES changes the default of 10 retries, and CLAUDE_CODE_RETRY_WATCHDOG=1 makes an unattended session retry 429 and 529 capacity errors indefinitely instead of failing (environment variables).

On the API, Anthropic's SDKs retry 5xx errors "twice by default" with exponential backoff, and they honor the retry-after header. Raise max_retries for batch work, then catch the 529 yourself: in Python, except anthropic.APIStatusError as e, check e.status_code == 529, and resend the same messages to the next model in your list. Copy one rule from Claude Code while you do: it never switches models on a rate limit, a billing or an authentication error, because another model can't fix your own account.

What is the 60 second fallback when every Claude model is overloaded?

Another vendor's model with your thread pasted in, which takes a minute only if the account already exists. Keep a one paragraph summary of the task (the goal, what's decided, the last thing that worked) in a note, paste it into GPT or Gemini, and carry on. The sibling guide what to use when ChatGPT is down lists the other providers and their status pages.

Whizi, which publishes this page, keeps Claude next to GPT and Gemini in one chat. Switching models in the middle of a conversation sends the existing thread to the newly selected model (how switching works), and a generation that fails is refunded, not charged. Where it loses: a message sent to Claude in Whizi still reaches Anthropic, so an incident across several models takes Whizi's Claude rows down too. The fix is the same switch to GPT or Gemini. Claude Sonnet needs Pro at $29.99 a month, Starter has no Claude, and Opus needs Powerhouse at $49.99. Every plan is capped at 60 messages an hour. With no developer API, Whizi replaces neither Claude Code nor your API fallback (Claude in Whizi).