Claude API Error 529 Overloaded: Live Status and What To Do

If Claude Code just printed API Error: 529 Overloaded and told you to check status.claude.com, that is not a billing bug and it is not your usage limit. It is Anthropic's overloaded_error: the API is at capacity across users. August 24, 2026 is a live incident. As of 06:42 UTC Anthropic still had an unresolved "Elevated errors for multiple models" event: claude.ai, the Claude API, Claude Code, and Claude Cowork were all on partial outage. Named models in the updates: Mythos 5, Fable 5, Opus 5, Opus 4.8, and "other Claude models." Console and Claude for Government stayed up.

This page is built to stay useful after today. The panel below reads Anthropic's public status API through explainx.ai's Is Claude down? checker, caches for about a minute, and refreshes on its own. When the incident closes, the same page still answers what a 529 is, how it differs from a 429, and what to do in Claude Code or your own API client.

TL;DR — is this my fault, and what do I do?

What is a Claude 529 Overloaded error?

Anthropic documents HTTP 529 as overloaded_error: the API is temporarily overloaded. The Claude Code CLI surfaces it in two shapes you will actually see:

API Error: 529 Overloaded. This is a server-side issue, usually temporary — try again in a moment. If it persists, check https://status.claude.com.

and, after the CLI has already retried for you:

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.

That second line is the important one. Claude Code retries 529s with exponential backoff before it gives up. If you are still staring at it, the platform — not your laptop — is the bottleneck.

Official API docs add the capacity detail: 529s happen when the API sees high traffic across all users. In rare cases a sharp traffic spike on your organization can also produce 429s from acceleration limits. Those are different codes on purpose. Ramp new production traffic gradually if you are bringing a new workload online; do not treat a fleet-wide 529 as a signal to fire more parallel jobs.

A 529 is the capacity-layer cousin of the July 29–30 network failures (elevated errors while traffic was rerouted) and a different failure mode from the August 16 authentication outage (you could not sign in). Overloaded means requests that do authenticate still get refused because serving capacity is gone.

529 vs 429 vs 500 — which error are you actually hitting?

The mix-up that wastes the most time: treating a 529 like a rate limit. People rotate API keys, upgrade plans, and restart Claude Code in a panic. None of that adds GPUs on Anthropic's side. If /usage still shows headroom and status.claude.com is yellow or red, stop spending the session on credentials.

The other mix-up is the reverse. The July 17 Fable access glitch showed a false "usage credits required" error while Fable was briefly unselectable — a product bug that looked like a limit. A vanished usage meter during an outage can look like "no limits" or "I'm out of quota" depending on which panel failed. Confirm against status.claude.com and /usage before you change plan or billing.

What is happening on August 24, 2026?

Treat this section as the dated snapshot. The widget above is the live answer.

Anthropic opened incident "Elevated errors for multiple models" at 05:06 UTC on August 24, 2026. The incident is still unresolved in the status feed as of 06:42 UTC. Four consumer/dev products flipped to partial outage; two stayed green:

That is a wide blast radius — web chat, the API, the CLI, and Cowork together — even though Anthropic's overall page indicator often reads "minor" while individual components sit on partial outage. If you are on Console or Claude for Government, this incident is not the one hitting you.

Timeline from Anthropic's own updates:

Anthropic has not published a root-cause postmortem in the incident feed. "Identified" means they think they know the fault, not that they have told the public what it is. That is the same disclosure pattern as the August 16 auth outage (~36 minutes, no public RCA) and the July 29–30 capacity incidents.

This is not a one-off — August has been a capacity month

status.claude.com's own past-incidents list for the two weeks before today is the useful context. Almost every weekday since August 12 has a degraded-performance or elevated-error row; Aug 21–23 were the only quiet days in that stretch:

Those are short incidents — most close the same day — stacked tightly enough that a 529 in Claude Code this month is more likely a platform event than a broken key. Trailing 90-day uptime in the mid-99s is compatible with that pattern: lots of brief partial outages, not a multi-day blackout. Plan for 529s as a normal operating condition at frontier-lab scale, not a surprise.

What should I do in Claude Code right now?

Claude Code already retries 529s for you — up to about ten attempts with backoff — before it prints the "Repeated 529" line. Fighting that with extra wrappers usually just adds load.

1. Look at the live panel on this page (or the full checker). If Claude Code is listed as partial outage, stop assuming your ANTHROPIC_API_KEY died.
2. Let the current retry cycle finish. Canceling and re-sending the same prompt in a tight loop is how a capacity incident gets worse.
3. Switch models with /model. Official Claude Code error docs are explicit: capacity is tracked per model, and the CLI will prompt you to switch when one model is under particularly high load (for example Opus). Sonnet or Haiku may still have headroom while Opus 5 or Fable 5 is saturated. Slash-command details live in the Claude Code commands reference.
4. Save state before you bail. Commit, or at least copy the current plan and file list, if you are going to jump to another harness. A recovered Claude session is useless if you lost the thread.
5. Fail over if you cannot wait. Keep a tested second path — GPT-5.6 Sol next to Claude Code, Cursor's router, or another CLI from the Claude Code vs Codex vs Gemini CLI comparison. An outage should cost a context switch, not an afternoon.

Do not rotate keys, do not "fix" billing, and do not paste the 529 into a new chat asking Claude to debug itself. The model cannot add serving capacity.

How should production API clients handle 529s?

Anthropic's official SDKs retry transient failures — connection errors, rate limits, and 5xx-class server errors — with exponential backoff, twice by default, and they honor a Retry-After header when present. You can raise that cap; you should not remove it and then invent a tighter loop.

from anthropic import Anthropic

# Default max_retries is 2. Raise it for batch jobs; keep it modest for
# interactive agents so a real outage fails fast enough to fail over.
client = Anthropic(max_retries=5)

try:
    message = client.messages.create(
        model="claude-sonnet-4-5",
        max_tokens=1024,
        messages=[{"role": "user", "content": "Summarize the diff."}],
    )
except Exception:
    # If this keeps throwing 529, check status.claude.com and fail over.
    # Do not spawn a second client that retries even harder.
    raise

If you own the retry loop yourself, the retry-with-backoff pattern is the whole policy:

• Retry 529, 429, and 500 with jittered exponential backoff.
• Stop after a small budget (for example 5 attempts or 30 seconds).
• If Retry-After is present, wait at least that long.
• Do not retry 400s, 401s, or 403s — those will not heal.
• When status.claude.com (or this page) still shows a partial outage after your budget, fail over instead of waiting forever.

A gateway that can shift Claude traffic to another provider is the grown-up version of /model. OmniRoute is the explainx.ai walkthrough for putting a free LLM proxy in front of Claude Code so a 529 degrades to a slower secondary instead of a hard stop.

Hammering a confirmed outage is the failure mode Anthropic is already under. Your retries are not free at their end.

When a 529 is not an outage

Three other things get mislabeled as "Claude is down":

You actually hit a 429. Weekly and 5-hour caps are still real in 2026 — they were not removed. If the error body says rate_limit_error, read the usage limits timeline instead of this page.

A usage panel fetch failed. During the August 16 auth incident, some Max users saw the weekly meter vanish and read that as "limits are gone." It was a broken fetch. A blank meter plus 529s is an outage symptom, not a policy change.

One model is hot, others are fine. Capacity is per model. Opus 5 can 529 while Sonnet still answers. Always try /model before you declare the whole platform dead.

If you want a second opinion besides Anthropic's own feed, the Is Claude down? tool also counts reader "report as down" clicks over 24 hours — a crowd signal that sometimes moves before the status page does, and sometimes lags it. Neither source is perfect; using both is better than using neither.

How to keep working when Claude is overloaded

Individual 529s last minutes. The August cluster is the reason to treat fallback as a default, not a drill.

• Interactive coding: second harness already authenticated, same repo, so you can switch mid-session. Compare options before you need them in the Claude Code vs Codex vs Gemini CLI guide.
• Production: health check the provider (this status API is https://status.claude.com/api/v2/status.json) before you decide to retry. A retry loop against indicator: major is wasted latency.
• Batch / async agents: a short queue absorbs a 30–60 minute incident with no user-visible failure. That is cheaper than paging a human for every 529.
• Alerts: subscribe on status.claude.com (email, SMS, Slack, Teams, webhook) once, so you are not manually refreshing during the next one.

The August 16 auth outage made the same point in 36 minutes. This 529 wave is the capacity-shaped version of that lesson.

Summary

A Claude API 529 Overloaded error is Anthropic's overloaded_error: shared capacity is full. It is not your rate limit, it does not burn quota, and retrying harder during a confirmed incident makes the queue worse. On August 24, 2026 that error matched a live partial outage on claude.ai, the Claude API, Claude Code, and Cowork, with Mythos 5, Fable 5, and Opus 5 named in the incident. Use the live panel on this page for current status; use /model, backoff, and a second provider for anything that cannot wait.

Related on explainx.ai

• Is Claude down? Live status checker
• Claude outage — authentication issue, August 16, 2026
• Claude outage — two network failures, July 29–30, 2026
• Claude usage limits — 2026 timeline explained
• Fable 5 access glitch — July 17, 2026
• OmniRoute — AI gateway / proxy for Claude Code
• GPT-5.6 Sol + Claude Code (Claudex) setup
• Cursor Router — auto model selection
• Claude Code vs Codex vs Gemini CLI
• Claude Code commands reference

Official sources: Claude Status · Anthropic API errors · Claude Code errors — 529 Overloaded

Live status is fetched from status.claude.com's public API and cached for about a minute. The August 24, 2026 incident timeline reflects Anthropic's status page as of 06:42 UTC that day. Always trust the widget and status.claude.com over any dated paragraph in this article.