Claude Overloaded Error (529): What It Means in claude.ai, Claude Code and the API

Claude’s “Overloaded” error, HTTP status 529 with the type overloaded_error, means Anthropic’s service is at capacity across all users. Anthropic’s API and Help Center pages call it temporary, and Claude Code’s reference says it is not your usage limit. claude.ai shows a related capacity message that Anthropic does not label as a 529. The documented response is to retry in a few minutes, check status.claude.com, and, in Claude Code, switch models with /model.

Key takeaways

• Three places, three wordings. The API returns a 529 overloaded_error, and Claude Code shows “API Error: Repeated 529 Overloaded errors”. claude.ai shows a related “capacity constraints” message that Anthropic does not label as a 529. Anthropic’s documents tie each to high demand across all users.
• It is not a limit. Claude Code’s documentation says a 529 does not count against your quota; Anthropic’s API pages do not say. The error for hitting a rate limit is a different one, 429.
• The documented fix is patience, then a different model. Wait a few minutes and check the status page. In Claude Code, /model switches to another model, because Anthropic says capacity is tracked per model. Claude Code retries transient failures up to 10 times before showing an error, so it has already retried by the time you see the message.
• A quiet status page does not rule it out. Anthropic’s Help Center says capacity constraints do not appear on the status page. In our saved copy of Anthropic’s incident feed (the 50 incidents it contained, July 27 to October 7, 2026), none of the 162 updates contains the words overload, capacity or 529. That is consistent with the Help Center’s statement but does not prove it.
• Two things Anthropic does not say. We found no Anthropic document that gives a specific duration or a cause for any particular overload.

Published October 12, 2026. We read Anthropic’s documents on October 12, 2026; the incident count uses the status feed we saved on October 10, 2026. We tested nothing ourselves: this article reports what the documents say.

What the Claude overloaded error means, in Anthropic’s own words

A 529 is an HTTP status code, and overloaded_error is the error type that comes with it. Anthropic’s API documentation defines it in one sentence: “The API is temporarily overloaded.” A warning under it adds that “529 errors can occur when the API experiences high traffic across all users.” The word “Overloaded” itself is the message text inside the error body; Anthropic’s streaming documentation shows {"type": "error", "error": {"type": "overloaded_error", "message": "Overloaded"}} (API errors, streaming, both retrieved October 12, 2026).

Claude Code’s documentation puts the same idea this way: the API is “temporarily at capacity across all users” (Claude Code error reference, retrieved October 12, 2026). In claude.ai, the Help Center uses different words for what reads as the same cause, high demand system-wide, but it does not say the chat message is a 529.

In claude.ai: the capacity constraints message

Anthropic’s Help Center article “Troubleshoot Claude error messages” has a section called Capacity constraints (retrieved October 12, 2026; the page shows a last-modified date of October 8, 2026). It says chat may answer with “Due to unexpected capacity constraints, Claude is unable to respond to your message”, followed by a request to try again soon.

The article is direct about what this is not: “Capacity constraints are not outages.” It describes the system as working normally while managing high demand across all users, calls the episodes temporary, and says they typically clear as demand shifts through the day, while advising a retry in a few minutes. Its advice is to try again in a few minutes. It also says capacity issues will not appear on the status page, because Anthropic treats them as normal load management rather than a technical problem.

Real service incidents are a separate category in the same article: disruptions where Claude is unavailable or significantly degraded for all or most users. Those are posted at status.claude.com. Usage-limit messages are separate too; they tell you when your limit resets, which a capacity message does not.

In the API: 529 and overloaded_error

For developers, the API errors page lists 529 overloaded_error next to the other 5xx codes. Four details matter:

• It is shared load. The page ties 529 to high traffic across all users.
• A sharp ramp-up can look different. In rare cases, a sudden increase in your own usage can produce 429 errors because of acceleration limits; Anthropic advises ramping up gradually.
• It can arrive mid-stream. When you stream responses, an error can occur after the API has already returned a 200, and the usual status-code handling does not catch it. The streaming page says an overloaded_error in the event stream normally corresponds to a 529 in a non-streaming call.
• The SDK retries by default. Anthropic’s official SDKs retry transient failures, including 5xx server errors, with exponential backoff, twice by default, and honor retry-after when it is present. The max_retries option changes this. Because 529 is a 5xx status, this appears to cover it, though the page does not name 529 in that sentence. This SDK default (twice) is separate from Claude Code’s own retry limit of 10 described below.

The page gives no 529-specific schedule. Some developers add their own backoff or reduce parallel requests; Anthropic’s documents do not prescribe this for 529, and we have not tested it. When you contact support, include the request_id from the error body.

In Claude Code: “Repeated 529 Overloaded errors”

Claude Code’s error reference has a section for exactly this message. By the time you see it, Claude Code has already retried: the reference says it retries transient failures up to 10 times with exponential backoff before showing an error. The 529 section says the API is at capacity across all users, that this is usually temporary, and that “A 529 is not your usage limit and doesn’t count against your quota.”

Its three recommended actions:

1. Check status.claude.com (or the provider’s status page named in the message) for capacity notices.
2. Try again in a few minutes.
3. Run /model and pick another model, “since capacity is tracked per model.” Claude Code may prompt this itself, for example “Opus is experiencing high load, please use /model to switch to Sonnet”.

A few related behaviors from the same documentation:

• Fallback chains. claude --fallback-model sonnet,haiku (or the fallbackModel setting) lets Claude Code switch automatically when the primary model is overloaded. The chain is capped at three models, applies to the current turn only, costs one turn with a cold prompt cache, and is unset by default.
• Slower retries. The variable CLAUDE_CODE_OVERLOADED_RETRY_BASE_DELAY_MS sets the starting backoff delay for 529 retries (default 500 ms, up to 32,000 ms; Claude Code v2.1.292 or later).
• Unattended runs. Setting CLAUDE_CODE_RETRY_WATCHDOG=1 makes sessions such as CI jobs keep retrying 429 and 529 errors indefinitely instead of stopping after the retry limit.
• Overload mid-response. If a server error arrives after Claude has finished a text block or tool call, Claude Code does not re-run the request, because that could execute the same tool calls twice. It keeps the completed work and shows “Server error mid-response. The response above may be incomplete.” (v2.1.199 or later). Our suggestion, not Anthropic’s: read what is already there before you resend.

How 529 differs from 429 and from an outage

Two other things get mistaken for an overload. The first is a rate limit. The second is an incident on the status page.

Claude Code also has a message, “Server is temporarily limiting requests (not your usage limit)”, which its reference describes as a short-lived throttle unrelated to your plan quota. The Agent SDK reference draws the same line in its error values: overloaded means a 529 because the server is at capacity, while rate_limit is a 429 against your quota.

The practical difference is the fix. Anthropic’s documents recommend waiting or switching models for a 529, and slowing your own requests or raising a limit for a 429. For the full five-step path from “is it just me?” to a decision, see our decision guide for when ChatGPT or Claude is not working.

What our incident data can and cannot say

Anthropic’s documents disagree slightly about the status page. The Help Center says capacity issues will not appear there. Claude Code’s reference tells you to check the status page for capacity notices. We cannot settle that from documents, so we checked the page’s own records.

We searched every incident title and update in the feed we saved on October 10, 2026 (50 incidents, 162 updates, started between July 27 and October 7, 2026), case-insensitive, for “overload”, “capacity”, “529” and “rate limit”. The result: zero incidents contain “overload”, “capacity” or “529”, in titles or updates. One contains “rate limit”: an October 7 incident about errors on the platform.claude.com Usage and Rate limits pages, a dashboard problem, not request throttling. A wider search for “demand”, “throttl”, “saturat”, “congest”, “too many” and “429” also returned nothing; “traffic” and “timeout” appear once, in a September 10 latency incident that attributes slowness to traffic entering through the US Midwest.

What this supports: in that window, Anthropic’s posted incidents did not use overload language. What it does not support: that no overload happened, or that incidents were never capacity-related. Titles are short and rarely explain a cause. Our study of all 50 Claude status incidents found that in only 3 of 50 incidents an update attributes the problem to a third party (by a three-phrase match) and that most describe what failed, not why. If an incident is posted, what “resolved” means is worth knowing before you decide how long to wait.

What we could not verify

• The word “Overloaded” in claude.ai. The Help Center documents the capacity constraints message, not the word “Overloaded”. We did not find the word “Overloaded” for claude.ai in any Anthropic document.
• How long an overload lasts. No Anthropic document we read gives a specific duration. The Help Center says episodes typically clear as demand shifts through the day and advises a retry in a few minutes; we do not state any other figure.
• Why a particular overload happened. The documents explain the condition, not individual events.
• Whether claude.ai’s message is a 529. Anthropic does not link the two in the pages we read.
• Per-model capacity details. “Tracked per model” comes from Claude Code’s reference; the API errors page does not repeat it.
• Anything about current fix success rates. We did not test retries, model switches or fallback chains. Claude Code behavior is described from the documentation on October 12, 2026 and may change with new versions.

Frequently asked questions