Corporate AI buyers can keep current DeepSeek integrations while adding a separately metered source of web search and page retrieval.
DeepSeek has paired an update to its Flash model with public Search and Fetch APIs, giving customers a way to bring current web material into applications through the company’s platform. In its October 17 DeepSeek announcement, the company said DeepSeek-V4.1-Flash-1017 was live across its website, app and API, and that both retrieval interfaces were open. The Decoder also reported that DeepSeek was making its in-house search engine available to outside developers.
The release separates a model update from new web-access services. Dealroom and MarkTechPost report that the updated model remains available under the existing deepseek-flash identifier. The revision’s Hugging Face model card lists an MIT license for the repository and model weights. The operational change is a familiar model name plus two paid interfaces: one for finding web results and one for retrieving a selected page.
Search and Fetch make web lookup two separately billed services
DeepSeek has added two billed services—one to find web results and one to return a specified page as Markdown—alongside the Flash update. Search accepts a query and returns structured web results. Fetch accepts one web-page address and returns the page’s content as Markdown, a plain-text format that preserves headings and other basic structure. The DeepSeek Search & Fetch guide says both are JSON-based web interfaces that use the same bearer credential as the company’s model API.
The documentation puts a clear commercial boundary around each service. Search and Fetch each cost US$0.75 per 1,000 successful requests. A Search request is billable only when it returns a successful response with results; a Fetch request is billable only when it produces page content. Failed requests, empty search results and timeouts are excluded under the posted terms. The number of results requested does not change the stated per-request price.
For companies using models to research, plan or generate content, the services offer a direct route from web lookup to model input through one provider. Their fees sit separately from model usage, so procurement teams can assess retrieval volume as its own line item. The source and quality of returned material remain governed by the search engine’s ranking and by the availability rules of the sites being fetched.
The stable model name limits immediate integration work
DeepSeek said the 1017 version brings stronger benchmark results, faster task completion and shorter outputs than the prior Flash release. It attributed the update’s project-building capability to reinforcement learning with self-feedback, a training method in which a model’s own evaluation signals help guide further training. The company also said it optimized the release for DeepSeek Harness, Tencent CodeBuddy and WorkBuddy.
The old public identifier gives existing users a smaller operational change to evaluate. Dealroom and MarkTechPost each describe the update as being served through deepseek-flash, so a customer can test the new version without first adopting a different model name. The newly introduced web endpoints still require their own requests, inputs and metering decisions.
DeepSeek’s benchmark table is useful context for the release, though it is company reporting. The model card lists a 42.3 score for V4.1-Flash-1017 on Terminal-Bench 4.0, compared with 31.2 for the preceding V4.1-Flash version. The Decoder said independent benchmark results had not yet appeared. That leaves buyers with a disclosed basis for comparison and a need to test results in their own workflows.
The two DeepSeek pages create a billing question to resolve
DeepSeek’s October 17 release post said Search and Fetch calls through server-side tools in DeepSeek Harness, Claude Code and Codex were free for API users until October 31 at 16:00 UTC. Its current guide describes a two-week fee-free period for Search requests made through Claude Code’s built-in Web Search tool or Codex’s web_search feature, then says the standard rate applies and a billing start date will be announced separately.
The two pages use different scopes and timing language. The published US$0.75 rate remains the relevant reference for regular Search and Fetch use, and the guide continues to specify the conditions that create a billable request. The difference makes current promotional eligibility a check to complete before a company commits a workflow to either channel.
The practical watchpoint is the billing notice or a revised guide that aligns the free-access terms across the tools DeepSeek named in its release. That clarification will determine when the new retrieval layer moves from a trial feature into a recurring operating cost alongside the stable deepseek-flash model identifier.