DeepSeek updated its V4.1 Flash model on October 17 and opened two separate web tools to API developers. The model now served under deepseek-flash is DeepSeek-V4.1-Flash-1017; the new Search endpoint returns structured web results, while Fetch returns one page as Markdown. For developers building an agent that needs current information, the release changes both the model available through an existing API name and the information sources an application can call. DeepSeek says the model also performs better on coding-agent benchmarks, but those figures are its own evaluations.

The big change

  • What changed: DeepSeek replaced the model behind deepseek-flash without changing that API name or its token prices. It also made Search and Fetch available as standalone JSON APIs.
  • Why it matters: An application can request search results, retrieve a selected page and provide that material to its own workflow. The documented interfaces make the steps and request charges visible; they do not establish that every retrieved page is accurate or that the new model performs better in a particular user's environment.
  • What to watch: Independent runs of the new model under disclosed agent settings would test DeepSeek's performance claims. Developers using Search or Fetch also need to track request billing, rate limits and how their applications handle page content that may be unavailable or untrustworthy.

One model name, a new version

DeepSeek's October 17 changelog says the updated Flash model is live on its website, app and API. Existing API calls using deepseek-flash reach V4.1-Flash-1017. The company says the calling method and model pricing are unchanged. Its current pricing table lists a one-million-token context length, support for thinking and non-thinking modes, tool calls and vision. It lists peak prices of $0.30 per million uncached input tokens and $1.20 per million output tokens; off-peak prices are half those rates. Cached input has a separate, lower rate. These are hosted-model charges, separate from Search and Fetch requests.

DeepSeek also posted the updated model weights under an MIT license. That gives teams a route to evaluate their own deployment, subject to the substantial compute and setup their chosen inference system requires. BIG CHANGE has not installed the weights or measured its resource needs. In its release post, the company recommends DeepSeek Harness v0.2.1-rc.1 or later for its latest model; that recommendation is for the harness, not an API version requirement.

The performance numbers are narrower than a universal coding claim. DeepSeek reports a DeepSWE v1.1 score of 77.0 for the new model versus 74.2 for its preceding V4.1 Flash, and a Terminal-Bench 4.0 score of 42.3 versus 31.2. Its changelog says the Terminal-Bench run used Claude Code and the DeepSWE run used Mini-SWE-V2. Those agent environments affect what a benchmark measures. The company also reports shorter task runtimes and outputs, but BIG CHANGE has not reproduced the tests or confirmed a cost-per-task improvement. None of these results establishes how the model will behave in an unattended production workflow.

Search results and fetched pages are separate calls

The Search and Fetch guide documents two POST endpoints: /v1/tools/search accepts a query and returns a results array, while /v1/tools/fetch accepts a URL and returns a title and page content as Markdown. Both require a DeepSeek API key in a bearer header. Search can request 1 to 50 results and optionally include page content; without that option, its content field is empty. Results include a URL, title, snippet, summary and metadata such as a date when available. These are fields supplied by the search service, not independent verification of a page's claims or publication time.

For an application that needs a current source, the documented path is to search for candidate pages and fetch a selected URL when its content is needed. The application still has to decide what to pass to the model and how to check the answer against the source. The guide does not say a bare deepseek-flash API call automatically performs those two requests. DeepSeek says server-side tools in DeepSeek Harness, Claude Code and Codex already call the services when used with its API; standalone integration remains a separate choice. This distinction matters when estimating both behavior and cost.

DeepSeek charges $0.75 per 1,000 successful Search requests and the same amount per 1,000 successful Fetch requests. A search must return a nonempty results array, or a fetch must return page content, to be billable under the published rule. The company says server-side Search and Fetch calls made through those named agent tools are free for DeepSeek API users until billing begins at 00:00 Beijing time on November 1, 2026 (16:00 UTC on October 31). The model's token charges still apply. Search and Fetch each have an account-level limit of five requests per second; Fetch also has a per-domain limit.

The Fetch endpoint depends on the target site's availability and access policies. DeepSeek's guide lists policy rejection, blocked or unsupported pages, no extractable content and rate limiting among possible failures. An application using the returned Markdown should treat it as source material to inspect, not instructions to obey. That is a workflow boundary for the developer, rather than a reliability guarantee from the API.

Sources & further reading