Anthropic has an active incident — elevated error rates are expected until it resolves.
A fix has been implemented and we are monitoring the results.
View Incident Detailarrow_forwardLearn more here: https://anthropic.com/news/fable-mythos-access
View Incident Detailarrow_forwardMost "Claude error today" reports come down to a handful of messages. Here is what each one actually means, whether it is your fault or Claude's, and the fastest way out.
{"type":"overloaded_error","message":"Overloaded"}What it means
Anthropic's inference servers are at capacity. HTTP status 529 is Anthropic-specific and means "too busy right now," not a bug in your request.
What to do
Retry with exponential backoff (wait 1s, 2s, 4s…). Switch to a lighter model (Sonnet/Haiku) if you were on Opus. During a real spike, even retries fail for a few minutes — that is normal.
This model isn't available right now. You can switch to another model to continue using Claude.What it means
A specific model tier (often the newest Opus) is capacity-constrained or mid-rollout. The rest of Claude is usually fine.
What to do
Do exactly what the message says — switch to another model (e.g. Sonnet) from the model picker and keep working. The original model typically returns within minutes to a few hours.
Spanner temporarily unavailableWhat it means
A transient hiccup in Anthropic's backend datastore. It surfaces as a failed request or a chat that won't send, and it clears on its own.
What to do
Wait 10–30 seconds and retry the same request. If it persists for more than a few minutes, check the live banner above — it usually means a broader incident is starting.
Claude is currently experiencing high demand. Please try again shortly.What it means
A softer form of the 529 capacity signal, shown in the web app rather than the raw API error.
What to do
Give it 30–60 seconds and resend. If you are on the API, treat it like a 529 and back off before retrying.
The page opens, you type a message, and nothing comes back (or it spins forever).What it means
The front door (the website edge) is up, but the inference layer behind it is failing. The page loading fast does not mean Claude can answer.
What to do
claude.ai can be up while the Claude API is down — check the Claude API status page if you build on the API. If everything looks green, hard-refresh and start a new conversation.
Building on the API? The HTTP status tells you instantly whether to retry or fix your request.
| Code | Type | What it means |
|---|---|---|
| 400 | invalid_request_error | Malformed request — your fault. Fix the payload. |
| 401 | authentication_error | Bad or missing API key. |
| 403 | permission_error | Key lacks access to that resource/model. |
| 429 | rate_limit_error | You hit your rate or token limit. Back off. |
| 500 | api_error | Unexpected server error on Anthropic's side. Retry. |
| 529 | overloaded_error | Anthropic is at capacity. Retry with backoff. |
Read the banner at the top of this page. Red means Anthropic has confirmed an active incident and errors like 529 Overloaded are expected. Amber means users are reporting problems but the official status page is still green. Green means no widespread errors — a one-off error you see is most likely local to your network or request.
HTTP 529 with an overloaded_error is Anthropic telling you its inference servers are temporarily at capacity. It is not a bug in your request. Retry with exponential backoff, or switch to a lighter model such as Sonnet or Haiku.
A specific model tier — usually the newest Opus — is capacity-constrained or mid-rollout while the rest of Claude works normally. Switch to another model from the picker to keep working; the original model typically returns within minutes to a few hours.
If the error is on Claude's side (529, 'overloaded', 'service is busy', 'model not available'), the fix is to wait and retry with backoff or switch models — you cannot fix capacity from your end. If it is a 400/401/403, the fix is in your request: correct the payload, API key, or permissions.
If the banner above is green and the live community feed on this page shows no spike in reports, the problem is most likely on your side — try a different network, hard-refresh, or a new conversation. If the banner is red or many users are reporting outages, it is Claude, and waiting is the only fix.