Anthropic API 529 "overloaded_error" in Claude Code: what it means, how long it lasts, and how to keep working

What it looks like, and the three errors it gets confused with

API Error: 529 {"type":"error","error":{"type":"overloaded_error","message":"Overloaded"},"request_id":"req_011CZ..."}

API Error: 500 {"type":"error","error":{"type":"api_error","message":"Internal server error"},"request_id":"req_011CZ..."}

API Error: 429 {"type":"error","error":{"type":"rate_limit_error","message":"This request would exceed your account's rate limit ..."}}
You've hit your limit · resets 4pm (Asia/Kuala_Lumpur)

Four different things print in the same place in Claude Code. They have different owners and different remedies:

Anthropic's own description of 529: "The API is temporarily overloaded. 529 errors can occur when the API experiences high traffic across all users." The same page notes that a sharp increase in your own usage can produce a 429 instead, from what it calls acceleration limits, and asks for gradual ramps.

How long it lasts

Anthropic publishes incidents at status.claude.com, and the GitHub threads link to them. The one behind the biggest Claude Code thread, Elevated error rates on Opus 4.6 on March 27, 2026, was opened at 06:59 UTC and resolved at 16:30 UTC, with a 0.4% error rate still reported as "high" in between. Others have been shorter: the April 8, 2026 incident lasted 19 minutes. There is no way to know from the error which kind you are in, which is why checking the status page first saves time.

Claude Code and the Anthropic SDKs retry 529 and 500 automatically with backoff before showing you the error. If the message reaches your terminal, those retries have already been spent.

Fix 1: change model, or wait

Overload is usually per model. /model in Claude Code switches the session to Sonnet or Haiku, which in the March incidents kept working while Opus did not. If every model is affected, the honest answer is to wait for the status page to turn green. Do not tighten a retry loop against a 529; the load is the cause.

Fix 2: if it is a 429 or a plan limit, it is not an outage

A rate_limit_error on an API key is your organization's tier: requests or tokens per minute, or the monthly spend cap. Per-minute limits clear on their own and the SDK backs off correctly; a spend cap has to be raised in the Console and returns no retry-after. The You've hit your limit · resets ... message is the Pro or Max subscription window. Several threads report it appearing with usage shown at 84%, and there is nothing to retry: the reset time is the reset time.

Fix 3: point Claude Code at another Anthropic-format endpoint

Claude Code speaks the Anthropic Messages API and nothing else, but it will speak it to any host, which is how it runs on Bedrock and Vertex. Our endpoint implements that API for open models, so the switch is a base URL, a key and a model name. Nothing is installed and nothing about the project changes.

export ANTHROPIC_BASE_URL="https://api.inferenceapis.com"
export ANTHROPIC_AUTH_TOKEN="your-inferenceapis-key"
export ANTHROPIC_MODEL="zai-org/GLM-5.3[1m]"                              # [1m]: tell Claude Code the window is 1M tokens
export ANTHROPIC_DEFAULT_HAIKU_MODEL="deepseek-ai/DeepSeek-V4.1-Flash[1m]"  # Claude Code's own background calls
claude
# to go back to Anthropic: unset the four variables, or open a new shell

For something you can flip per project instead of per shell, put the same variables in the project's .claude/settings.json. Remove the block, or move it to a file you keep for outages, to return to Anthropic.

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.inferenceapis.com",
    "ANTHROPIC_AUTH_TOKEN": "your-inferenceapis-key",
    "ANTHROPIC_MODEL": "zai-org/GLM-5.3[1m]",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek-ai/DeepSeek-V4.1-Flash[1m]",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "zai-org/GLM-5.3[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek-ai/DeepSeek-V4-Pro[1m]",
    "CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC": "1"
  }
}

We ran this on September 21, 2026 with Claude Code 2.1.278 in headless mode: given a file with a bug and a failing test, GLM 5.3 through this endpoint ran the test, fixed the bug, re-ran it and reported done in 6 turns and 3 seconds. Claude Code prints [claude-code:unrecognized_model] on stderr for a model name it does not know; it is a note about its own price table and the run continues. Two more details from the full Claude Code guide: the [1m] suffix tells Claude Code the model's window is 1M tokens so it does not compact early, and CLAUDE_CODE_DISABLE_THINKING=1 switches the model's reasoning off for quicker routine edits.

Which model, and what it costs

None of these is Claude, and for hard problems you will feel the difference. For getting through an outage, or a locked plan window, they finish real tasks. From our test run of the same fix-the-tests task on every model (September 18, 2026), billed cost from our own logs:

A Max plan window that has locked you out costs nothing more to wait out; these cost cents per task and stop when you stop. Prices are flat, hosting is in the United States and nothing is stored after the response (/trust).

Frequently asked questions

Is 529 a real HTTP status code?

It is not in the HTTP standard. Anthropic uses it for overload so that it is distinguishable from 503 and from its own 500 api_error. Generic HTTP clients treat it as a server error, and the Anthropic SDKs retry it.

Does a 529 count against my usage or bill?

No. The request was not served. Retries that eventually succeed are billed once, as normal.

Will my Anthropic prompt cache survive switching endpoints and back?

No. Anthropic's cache is per organization and expires within minutes anyway; after a switch the first request rebuilds it. Our endpoint bills repeated prompt prefixes at a cached-input rate automatically, so the same effect exists on our side.

Can I keep Claude for the main model and use this only for background calls?

Not from one Claude Code process: ANTHROPIC_BASE_URL applies to every request it makes. You can run one project on each by putting the variables in that project's .claude/settings.json.

Does thinking work on these models through Claude Code?

Yes. Claude Code sends thinking: {type: "adaptive"}; the endpoint accepts enabled, adaptive and auto, keeps the model's reasoning on, and streams it as thinking blocks. The 400 "type must be in [enabled, disabled, auto]" that some hosts return does not occur here.

Where Inference APIs fits

Fix 3 is what we sell: Claude Code talks to any endpoint that speaks the Anthropic Messages API, and ours does, for DeepSeek V4, GLM 5.3, Kimi K2.7, MiniMax M3 and GPT-OSS. None of them is Claude. They did complete our Claude Code test task end to end, for cents, and they are there when Anthropic is not. Models run on GPUs in the United States with zero retention, billed per token from a prepaid balance; a new account starts with $1 of credit.

Related entries

• Claude Code on a third-party model: API Error 400 'type' must be in ["enabled", "disabled", "auto"]
• Claude API 404 not_found_error: retired model ids (claude-3-5-sonnet, claude-3-7-sonnet, claude-sonnet-4, claude-opus-4)
• DeepSeek API 429 "Rate Limit Reached" and 503 "Server Overloaded": the concurrency cap behind both, and your options
• HTTP 429 rate_limit_exceeded (Groq)