Which models DeepSeek has right now
As of September 2026, DeepSeek has two current API identifiers. deepseek-v4-pro (DeepSeek-V4-Pro-0813) became generally available on 13 August 2026 and is the heavy model for hard reasoning. deepseek-flash (DeepSeek-V4.1-Flash) appeared on 10 September 2026: it is the smallest model of the new architecture, with native understanding of multimodal input.
The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp from earlier releases are kept for backward compatibility, and requests to them are routed automatically to the new deepseek-flash. If one of these identifiers shows up in your error logs, you are in effect dealing with the same new flash, not with a separate retired model.
The service for V4-Pro was first due to be wound down on 14 September 2026, but DeepSeek extended access past that date without changing pricing, according to the official changelog at api-docs.deepseek.com/updates.
The 14 September outage: what actually happened
On 14 September 2026, the same day as the original V4-Pro shutdown date, DeepSeek's chat and API suffered a mass outage. According to a third-party write-up, up to 438 user reports of unavailability arrived within an hour, while DeepSeek's official status page showed "all systems operational" the whole time (startupfortune.com).
It is a textbook case of a status page diverging from what happens to real requests. A provider's automatic monitoring often checks only that an endpoint is reachable, not whether it returns real answers to application requests. A server can answer 200 to a health check and still time out or return empty responses to genuine model calls.
At the time of writing, 25 September 2026, status.deepseek.com again shows every service as operational. Check the incident history in the same place, under History.
What to do if DeepSeek doesn't respond right now
Check whether the outage is real
Before changing code, rule out local causes: a wrong key, an empty balance, a requests-per-second limit exceeded. Authorization and rate-limit errors look different from timeouts and dropped connections, and you fix them without waiting.
Don't take the status page at its word
The 14 September story shows that a green status does not guarantee that requests go through. Several requests in a row time out or return an empty response? Don't wait for the provider to update the page; switch to a fallback path.
Keep a plan B on another provider
The only reliable protection against one API's downtime is the ability to move quickly to another. That requires a client that is not tied to a specific provider. When one base_url and one key format are hard-coded, switching takes hours instead of minutes.
DeepSeek and a backup on other models through codexkey
codexkey carries DeepSeek: deepseek-v4-pro ($0.12 / $0.34 per 1M tokens, input / output), deepseek-v4-flash ($0.08 / $0.22) and deepseek-v4.1-flash ($0.04 / $0.16), per the price list of 29 October 2026. You pay with a Russian bank card, and no VPN is needed. We cannot promise that a DeepSeek outage will not touch these models too, so keep a model from another vendor in reserve. The same codexkey key gives access to Claude, GPT and Gemini, and they can cover an urgent need for a working API while you sort out the main provider.
The API is compatible with the Anthropic SDK, and the configuration changes through two variables:
export ANTHROPIC_BASE_URL="https://api.codexkey.online"
export ANTHROPIC_API_KEY="sk_cdx_your_key"
In Python, the client takes base_url explicitly:
import anthropic
client = anthropic.Anthropic(
api_key="sk_cdx_your_key",
base_url="https://api.codexkey.online",
)
The takeaway for anyone building a production pipeline on an LLM API: make the code provider-independent from day one, not after the first outage. One environment variable in place of a hard-wired client saves hours exactly when the main provider is down. The list of models and current prices are on the models page.
FAQ
Why does DeepSeek V4 Pro return an error when the status page is green?
Because the status page's automatic monitoring does not always reflect how requests are really processed. That divergence was confirmed on 14 September 2026: hundreds of outage complaints while the status stayed green. Trust your own error logs first, and use the official page as a cross-check.
What happened to DeepSeek on 14 September 2026?
A mass outage of the chat and API, recorded by third-party observers from the number of user complaints within an hour, while the official status page did not show the incident. DeepSeek has not made the technical cause public.
Are DeepSeek Flash and V4 Pro the same thing?
No. deepseek-flash is the smaller, cheaper model of the new architecture, released on 10 September 2026. deepseek-v4-pro is the flagship for hard tasks, generally available since 13 August 2026. Old identifiers such as deepseek-v4-flash are routed to the new flash.
Will DeepSeek V4 Pro be shut down?
The service was originally to end on 14 September 2026, but DeepSeek extended access without changing rates, judging by the official changelog. No other shutdown plans had been announced as of September 2026.
How do I protect myself from another DeepSeek outage?
Keep a backup API provider on different infrastructure and don't hard-wire your code to one base_url. Then switching during an outage takes minutes, and you don't have to rewrite the integration on the fly.