DeepSeek ships a faster V4.1 Flash update and opens a paid web-search API, two days after a nearly five-hour outage

DeepSeek released an updated version of its low-cost V4.1 Flash AI model on Oct. 17 and, on the same day, began selling outside developers access to the web search engine behind its own chatbot. Both moves came two days after the Chinese company's chatbot and main developer services were disrupted for almost five hours.

What DeepSeek released

DeepSeek said on Saturday, Oct. 17, that it had replaced its V4.1 Flash model with an updated version, DeepSeek-V4.1-Flash-1017, across its website, app and developer API. Developers keep calling the same model name, and the company said API pricing is unchanged.

In its technical notes on Hugging Face, DeepSeek calls the release a minor update. It says the model scores better on reasoning, coding, agent and image tests while finishing tasks faster and writing shorter answers, which the company says lowers the cost of each task. The weights are free to download under the MIT License.

The gains are DeepSeek's own test results. On Humanity's Last Exam, a broad reasoning test run without tools, the company reports a score of 42.6, up from 36.8 for the September version. On Terminal-Bench 4.0, a test of work inside a computer terminal, it reports 42.3, up from 31.2.

DeepSeek's own comparison table also shows where the model still trails. Anthropic's Opus 5 is listed at 56.3 on the reasoning test and 51.8 on the terminal test. OpenAI's GPT-5.6 Sol is listed ahead on the reasoning test at 44.5 but behind on the terminal test at 39.9. All of these figures come from DeepSeek's own testing.

A new paid product: web search

The same day, DeepSeek opened two new paid services to outside developers. A Search API returns structured web results from the in-house search engine DeepSeek already uses for its own chatbot, and a Fetch API returns the content of a single web page. DeepSeek says its engine handles hundreds of millions of search requests a day and indexes hundreds of billions of pages.

Each service costs $0.75 per 1,000 successful requests. Searches and page fetches made through coding tools such as DeepSeek Harness and Claude Code are free for now, and per-request billing for that route begins at midnight Beijing time on Nov. 1.

That gives DeepSeek something to sell besides access to its models. The model prices themselves stay low: the Flash model costs $0.30 per million uncached input tokens and $1.20 per million output tokens at peak hours, and half that off-peak.

The outage two days earlier

The update came two days after a disruption to DeepSeek's chatbot and its main APIs. DeepSeek's status page shows the incident ran from 8:11 a.m. to 1:04 p.m. Beijing time on Thursday, Oct. 15, a total of 4 hours and 52 minutes, and affected both its V4 Pro and V4.1 Flash APIs as well as chat.

DeepSeek said it had found the cause and fixed it, but did not say what went wrong. Caixin counted it as the company's fifth outage in October. The South China Morning Post reported that DeepSeek apologized and offered users credits refunding what they spent from Oct. 8 to Oct. 14. Some industry observers have suggested the company is short of computing power, which DeepSeek has not confirmed.

Why markets will watch

DeepSeek is private, but its releases have moved listed rivals. After it launched the original V4.1 Flash on Sept. 10 at a steep discount to rivals, shares of Hong Kong-listed MiniMax and Z.ai fell more than 8%.

This update is smaller and comes without a price cut, so a repeat is not a given. Monday, Oct. 19, is a public holiday in Hong Kong, so Tuesday, Oct. 20, is the first trading day there since the release. The bigger open question is whether DeepSeek can keep its service reliable while it adds new paid products.