DeepSeek not working: how to check status and fix it

When DeepSeek stops responding, open status.deepseek.com and look at the indicator for the service you use. The chat and the API are listed separately, so you can tell in seconds which half is down. Is the page green while your requests fail? Then the cause is usually your key, your balance, your request rate or your network. Not always, though: the page can lag behind reality, and there is an example below. The check takes half a minute. Sorting causes by symptom takes a bit longer.

How to check DeepSeek status in 30 seconds

1. Open status.deepseek.com. The top line reads "Everything is running smoothly" or lists problems. Below it come separate indicators: DeepSeek V4 Pro API, DeepSeek V4.1 Flash API, Search API and Fetch API, the chat (Chat Service), file upload and search. The chat, the API and search fail independently of each other.
2. If the complaint is fresh, click "View historical events" under the calendar. It shows whether anything was logged in the last few hours.
3. Subscribe if your work depends on DeepSeek. The page has a "Subscribe to updates" button, plus feeds at status.deepseek.com/feed.rss and status.deepseek.com/feed.atom. A subscription saves you from checking by hand.
4. Send one minimal request from a different network, for example mobile data instead of home Wi-Fi. If it goes through, the problem is your connection or your client.
5. Status green but requests time out or come back empty? Do not wait for the page to catch up. Look at your own logs and the response code.

The indicator list, chat and API limits, and installing the app without fakes are covered in more depth in DeepSeek for free: chat, API limits and service status. This article does not repeat them.

Why DeepSeek is not working: causes by symptom

The codes and wording in the table come from the error page of the DeepSeek docs. The first three rows are the server's side. The rest are your settings or your network. Retrying 401, 402, 400 and 422 is pointless, because they fail the same way every time.

Region and VPN

We found no official list of countries where DeepSeek is unavailable. In DeepSeek API from restricted regions we describe how the site and the API platform open from Russia without a VPN, and how the trouble starts at payment. If the chat will not load, test the same request with a VPN and without one. Both take a minute, and one of them may be the cause. Errors 401, 402 and 429 have nothing to do with your region.

What to do in the chat and the app

1. Check the chat indicator (Chat Service) on status.deepseek.com. If an entry is open, wait: reloading will not help.
2. Status green? Refresh the page, start a new conversation with a short message such as "hi" and see whether the service answers at all.
3. Switch the network and the browser or app. Mobile data and a different browser rule out half of the local causes.
4. Check the neighbouring services. If the chat answers but search or file upload does not, look at their own rows, because they have their own indicators.
5. Install the app only from the download page on deepseek.com. A "DeepSeek client" from anywhere else may be someone else's build.

What to do in the API

Start with the response code and the error text. A 400, 401, 402 or 422 points at your request, key or balance, and sending the same request again will fail the same way. A 429 is about rate: after a pause or a freed slot the same request can go through. A 500 or 503 is temporary according to the docs: retry with a growing pause and a random spread, though success is not guaranteed while the overload lasts. Ready-made exponential backoff with jitter for Python and TypeScript is in the Claude API errors guide, and the logic for DeepSeek is the same.

Three details that eat time:

• The concurrency limit is counted per account, not per key, so extra keys do not raise it. While a request waits in the queue, the server keeps the connection open and sends empty lines or SSE comments : keep-alive. That is not a failure, but a hand-written response parser can trip on them. If output has not started within 10 minutes, the connection is closed.
• The current model IDs are deepseek-flash and deepseek-v4-pro. The old deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted, but those models are retired: requests go to V4.1 Flash and are billed at the Flash price.
• Key and balance take a minute to check: a wrong key gives 401 and an empty balance gives 402.

A minimal test request, in the format from the DeepSeek docs:

curl https://api.deepseek.com/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer ${DEEPSEEK_API_KEY}" \
  -d '{"model": "deepseek-flash", "messages": [{"role": "user", "content": "Answer in one word: are you up?"}], "stream": false}'

If a reply comes back, the service is alive and the cause is in your client. No reply from a second network either? Open status.deepseek.com and the feed.

The 14 September 2026 outage and a green status page

On 14 September 2026 users reported that the DeepSeek chat and API were down. A third-party write-up on startupfortune.com cites Entireweb Status data: 491 reports in 24 hours, 438 of them in a single hour, while the status page showed every service as operational. DeepSeek has not publicly explained the cause.

Today the status feed contains an entry titled "Web/API Partially Unavailable" dated 15 September, 02:30 Beijing time, which is 18:30 UTC on 14 September. It lists the chat and the V4.1 Flash API and is closed. The feed does not show when the entry first appeared on the page. The practical lesson: "the service is reachable" and "the model answers your request" are different things, so keep your own error logs and do not rely on one green indicator.

Status as of 10 November 2026

We opened the status page, its feed and the DeepSeek docs. We did not test the chat or the API with requests.

• The status page shows the top line "Everything is running smoothly" and no open incident.
• The feed holds 24 entries as of 10 November, from 14 August to 7 November 2026, all closed: 15 degraded-performance entries, 6 partial outages and 3 search failures. That averages about two entries a week, so subscribing to the feed pays off.
• Uptime for August to November 2026, as calculated by the page itself: V4 Pro API 99.74%, V4.1 Flash API 99.56%, chat group 99.82%.
• Models: deepseek-flash (DeepSeek-V4.1-Flash-1017, updated on 17 October, prices unchanged) and deepseek-v4-pro (DeepSeek-V4-Pro-0813). In its 10 September announcement DeepSeek wrote that it was winding down V4 Pro and would route its requests to V4.1 Flash from 14 September. The change log says otherwise: the V4 Pro API keeps running after 14 September on the same billing, and the price list on 10 November keeps it as a separate model with its own prices. The change log has no newer announcement on the subject.

The latest feed entries (entry time in UTC; the feed itself stamps them in Beijing time):

A backup route through Claudexia

DeepSeek is available in Claudexia: deepseek-v4-pro ($0.06 / $0.18 per 1M tokens, input / output), deepseek-v4-flash ($0.12 / $0.36) and deepseek-v4.1-flash ($0.06 / $0.22), from the public Claudexia catalog, Premium channel, on 10 November 2026. Prices change, so check the models page. You can top up by card, by SBP bank transfer in Russia, or with USDT, BTC or ETH, and no VPN is needed. We cannot promise that an outage at DeepSeek will not touch these models too, so keep a model from another vendor in reserve. Claude, GPT and Gemini are reachable through the same Claudexia key, which can cover an urgent need for a working API while you sort out your main provider. The live status of the Claudexia API and models (30-day uptime, time to first token, incidents) is published on the status page.

The API is compatible with the Anthropic SDK, and configuration changes with two variables:

export ANTHROPIC_BASE_URL="https://api.claudexia.tech"
export ANTHROPIC_API_KEY="sk_cdx_YOUR_KEY"

For Python the client takes base_url explicitly:

import anthropic

client = anthropic.Anthropic(
    api_key="sk_cdx_YOUR_KEY",
    base_url="https://api.claudexia.tech",
)

The takeaway for anyone building a production pipeline on an LLM API: keep the code provider-agnostic from the start, not after the first outage. One environment variable instead of a hard-coded client saves hours at exactly the moment the main provider is down. The model list and current prices are on the models page.

FAQ

Why is DeepSeek not working today?

There are three causes. An outage at DeepSeek, visible on status.deepseek.com and in the feed. Your settings: a wrong key (401), an empty balance (402), requests that are too frequent or over the concurrency limit (429). Your network or client, in which case the request goes through from another connection. The order of checks is above, and the whole thing takes about a minute.

How do I check DeepSeek status?

Open status.deepseek.com. It has separate indicators for the V4 Pro API, the V4.1 Flash API, the chat, file upload, search, the Search API and the Fetch API, plus a subscription and RSS and Atom feeds. The history of entries is behind the "View historical events" link.

Why does DeepSeek return errors when the status page is green?

The indicator can lag, and the error can be yours. 401, 402, 429, 400 and 422 have nothing to do with server health. If requests time out or return empty replies while the status is green, trust your own logs: on 14 September 2026 the complaints came precisely while the page was green.

What do 503 and 429 mean in the DeepSeek API?

According to the docs, 503 means the server is overloaded by high traffic, and DeepSeek advises retrying after a short pause. While the overload lasts, a retry may still fail. 429 means you are sending requests too fast or have exceeded the concurrency limit. The pause and the queue are yours to add, not DeepSeek's.

Are DeepSeek Flash and V4 Pro the same model?

No. deepseek-flash is the smaller, cheaper model of the new architecture, released on 10 September 2026 and updated to the 1017 version on 17 October. deepseek-v4-pro is the heavy model for complex tasks, generally available since 13 August 2026. Old IDs such as deepseek-v4-flash are routed by DeepSeek to the new flash.

Will DeepSeek V4 Pro be shut down?

It was first slated to wind down on 14 September 2026. DeepSeek changed course: its change log says the V4 Pro API keeps running after that date on the same billing. As of 10 November 2026 we found no newer shutdown announcement, but plans have already changed once, so keep an eye on api-docs.deepseek.com/updates.

How do I protect myself from the next DeepSeek outage?

Subscribe to the status feed, keep a backup API provider on different infrastructure, and do not hard-code a single base_url. Then switching during an outage takes minutes instead of a rewrite of the integration on the fly.