Claude going quiet mid-conversation, the app showing an error, or your API integration suddenly returning 5xx — these all feel identical from where you sit, but they have different causes and different fixes. Anthropic can be fully down, degraded on one surface (the API but not the web app, or one model but not another), rate-limiting your specific key, or completely healthy while your own network is the thing failing. Here's how to tell which, quickly.
The surfaces that fail independently ¶
"Claude" is really several products sharing infrastructure, and Anthropic's status page tracks them separately:
- claude.ai — the consumer web app and its login/session layer.
- The API (api.anthropic.com) — what your code, your IDE assistant, and third-party tools call.
- Per-model availability — a specific model family can be capacity-constrained or degraded while others respond normally, especially right after a new model ships.
- Console / billing — the dashboard where you manage keys and usage can have its own hiccups without the API being affected.
The practical upshot: "Claude is down" is rarely all-or-nothing. Pin down which surface before you decide it's an outage.
The under-a-minute check ¶
- Probe Anthropic from four regions. Check claude.ai and api.anthropic.com from US East, US West, Europe, and Asia. All four failing on both is a global outage; the API failing while the app passes (or the reverse) tells you it's surface-specific.
- Read the incident feed. Our Anthropic status page mirrors Anthropic's official feed, so you can see whether they've acknowledged an incident and which component it hits.
- Compare against the field. The AI provider status board shows Anthropic next to OpenAI, Google, Groq, and Mistral — the fastest way to tell "Claude is down" apart from a broader AI-infrastructure event. If ChatGPT and Claude are both erroring at once, suspect a shared upstream rather than either provider.
Outage vs. rate limit vs. auth ¶
For the API, the HTTP status code is the diagnosis. Don't guess — read it:
| Status | Meaning | Is Anthropic "down"? |
|---|---|---|
| 500 / 503 / 529 | Server error / overloaded | Yes — real outage or capacity event. Retry with backoff. |
| 429 | Rate limit | No. Your key hit an RPM/TPM/daily cap. Back off → |
| 401 / 403 | Auth / permission | No. Bad, revoked, or out-of-credit key. |
| 400 / 422 | Bad request | No. Your payload — usually context-length or a malformed field. |
Anthropic returns an "overloaded" signal (HTTP 529) during capacity crunches. It looks alarming but is transient — treat it exactly like a 503: retry with exponential backoff, don't hammer.
When the check says it's you ¶
If all four regions reach Anthropic but you can't, the fault is local. Common causes: a corporate proxy or firewall breaking the streaming (SSE) connection Claude uses for responses, a VPN interfering, stale DNS, or an out-of-credit/expired API key that looks like an outage but is really a 401. Fast tests: open claude.ai in a private browser window, try a different network (mobile hotspot), and confirm your key still has credit in the console.
Building on the API? Fail gracefully ¶
If you ship a product on Claude, a brief provider blip shouldn't take your app down with it. Treat 429 and 5xx (including 529) as retryable with exponential backoff and jitter, cap the retries, and surface a friendly "the AI is briefly unavailable" state instead of a stack trace. A short Anthropic incident then becomes a few seconds of degraded UX rather than an outage of your own.
FAQ ¶
Is Claude down for everyone or just me?
Probe claude.ai and api.anthropic.com from multiple regions. If every region fails, it's down for everyone. If they all succeed but you can't connect, it's your network, proxy, or key. If only one surface (say the API) fails while the app works, it's a component-level issue, not a full outage.
What does an "overloaded" (HTTP 529) error from Anthropic mean?
It means Anthropic's servers are temporarily at capacity and are shedding load — the service is up but can't take your request right now. It's transient; retry with exponential backoff rather than immediately re-sending. It is not a permanent outage and usually clears within seconds to minutes.
Claude works in the browser but my API calls fail — why?
The web app and the API are separate surfaces on shared infrastructure and can fail independently. If claude.ai is fine but your integration errors, probe api.anthropic.com directly and read the status code: a 5xx is an API-side problem, a 429 is a rate limit, and a 401 is an auth issue on your key.