DeepSeek has released V4.1-Flash-1017, an open-weight model update it says is faster, more capable on coding and agent work, and more economical than its predecessor. At the same time, the company has opened Search and Fetch APIs that let developers query the web and retrieve page content, with usage-based pricing and a limited free period. The release is significant as a product and infrastructure change, but its performance comparisons remain company-reported rather than independently established.

DeepSeek pairs V4.1-Flash-1017 with newly open web-search APIs

The release in context

DeepSeek announced V4.1-Flash-1017 on October 17, 2026, saying that it was immediately available on the company’s website, app, and API. The announcement also links open-weight files on Hugging Face and ModelScope. DeepSeek’s release notice describes the update as a Pareto improvement: better performance and faster completion while producing shorter outputs and lowering cost, according to the company.

The announcement presents the model as especially suited to end-to-end software work. DeepSeek says reinforcement learning with self-feedback, referred to in the company’s material as RLSF, is intended to improve tasks such as building games, creating 3D models, and developing complete projects. It also names DeepSeek Harness, Tencent CodeBuddy, and WorkBuddy as development environments for which the model was optimized, and recommends DeepSeek Harness v0.2.1-rc.1 or later for the best experience. Those are product and training descriptions from DeepSeek, not findings independently reproduced in the supplied reporting.

This is a release-and-infrastructure story rather than a funding story. The practical question is not who financed DeepSeek, but how a model update and a web-access layer might change the way developers assemble research, retrieval, and coding agents. That distinction matters because a model can be inexpensive or impressive on a published chart without automatically being reliable in a production workflow. The announcement establishes availability and pricing; it does not by itself establish broad superiority.

What Search and Fetch do

The new tools are two separate raw JSON APIs. The official guide says Search finds web pages for a query and returns structured results, while Fetch retrieves one web page and returns its content as Markdown. Together, they are aimed at agents and retrieval-augmented generation applications that need current web material rather than only information contained in a model’s training data.

Both services use the base URL https://api.deepseek.com and accept POST requests. Search is available at /v1/tools/search; Fetch is at /v1/tools/fetch. Requests use the same bearer authentication as the DeepSeek API, with an Authorization: Bearer <DeepSeek API Key> header and JSON content type. Search accepts a required query, an optional result count from 1 to 50, an optional original question, localization data, and a need_content flag. Fetch accepts a URL and returns a page title and Markdown content when the target can be retrieved and extracted.

DeepSeek’s guide sets an account-level limit of five queries per second for Search and five per second for Fetch; the limits are independent. Fetch also has a per-domain limit. A request above the account limit receives HTTP 429, while other documented responses cover invalid credentials, insufficient balance, blocked or unavailable pages, and timeouts. These details make the APIs more concrete than a generic “web access” feature: developers must manage authentication, retries, target-site availability, and the possibility that a page has no extractable content.

The listed price is $0.75 per 1,000 successful Search requests and $0.75 per 1,000 successful Fetch requests. The unit is a successful request, not a token or the number of results returned. The guide says an empty Search result, failed request, or timeout is not billed; Search must return a non-empty results array, and Fetch must return page content. DeepSeek’s announcement says Search and Fetch calls made through specified server-side tools in DeepSeek Harness, Claude Code, and Codex are free for DeepSeek API users until October 31, 2026, at 16:00 UTC. That free period applies to the stated eligible use, not necessarily to every possible API call.

What the benchmark evidence shows

DeepSeek’s announcement claims stronger benchmark performance, quicker task completion, shorter outputs, and improved coding and agent performance. The supplied independent coverage, The Decoder’s report, reproduces readings from DeepSeek’s own charts. It reports a 77% DeepSWE v1.1 result at just under $0.20 per task, a lower setting above 72% at about $0.10, and a predecessor result of 74.2% at around $0.50 per run. The report also describes a DeepSeek chart placing several rival models at roughly $2 to more than $10 per agentic coding task.

Those figures should be read as chart readings, not as an independent study. The Decoder explicitly says independent benchmarks were not yet available. Neither the company’s benchmark chart nor a report that summarizes it supplies independent confirmation of the test design, implementation, or real-world code quality. The safe conclusion is narrower: DeepSeek reports an efficiency and performance improvement, and an independent publication has documented what the company’s charts claim. It is not safe to state as fact that V4.1-Flash-1017 beats larger rivals.

Benchmark context also matters. A Pass@1 score and a per-task cost can reveal useful trade-offs, but they do not answer every production question. They do not, on the supplied evidence, establish how often an agent makes subtle errors, how much human review is required, how robust it is to changing repositories, or how well its web retrieval handles difficult sites. The API guide’s own mechanics underline this point: Fetch depends on the target site’s availability and access policies, and a successful HTTP transaction is not the same thing as a trustworthy answer.

Why the combination matters

Model releases and retrieval APIs are often discussed separately, but this announcement connects them. An agent that can write or modify code may also need to search for documentation, inspect a live page, compare implementation guidance, or ground an answer in information that changed after training. Search supplies candidate pages; Fetch supplies a machine-readable page body. That division can make a workflow easier to assemble than one that requires a developer to build separate discovery and extraction services.

The economic design is also easy to understand. At $0.75 per 1,000 successful calls, a team can estimate retrieval cost separately from model-token cost. Because the price does not rise with the number of returned Search results, the count parameter changes response breadth rather than the listed per-request fee. But the five-QPS account limits and Fetch’s domain limit put a ceiling on bursty workloads. A production system would still need caching, backoff, content-quality checks, and controls for duplicate or low-value retrieval.

Open-weight availability adds another route for experimentation, because the model files are linked for download on two model repositories. The supplied facts do not establish a particular license interpretation, hosting cost, hardware requirement, or suitability for a given organization, so those questions should not be inferred from the phrase “open-weight.” Likewise, the API release does not establish that every webpage is accessible or that every result is accurate.

For developers, the most defensible near-term reading is practical rather than triumphalist: there is a newly released model to test, a documented web-search and page-fetch interface to integrate, a transparent request price, and a temporary waiver in specified server-side contexts. The performance case remains provisional until independent evaluations and hands-on production experience are available. This article describes the release; it is not investment or legal advice.

FAQs

What is DeepSeek V4.1-Flash-1017?

It is DeepSeek’s model update announced on October 17, 2026. DeepSeek says it is live on its website, app, and API, and links open-weight files on Hugging Face and ModelScope. The company positions it for coding, agent tasks, and end-to-end project development.

What are the Search and Fetch endpoints?

They are raw JSON POST APIs at https://api.deepseek.com/v1/tools/search and https://api.deepseek.com/v1/tools/fetch. Search returns structured web results; Fetch retrieves one page as Markdown. Both use bearer authentication.

How much do the APIs cost?

Each endpoint costs $0.75 per 1,000 successful requests. The official guide says billing is per successful request rather than tokens or result count, and that empty searches, failed requests, and timeouts are not billed.

Are the claimed benchmark wins independently verified?

Not in the supplied coverage. The Decoder says independent benchmarks were not yet available. Its reported scores and cost figures are readings of DeepSeek’s internal charts, so comparisons with rival models must remain attributed to DeepSeek.

When is the free period?

DeepSeek says eligible API users can use Search and Fetch through specified server-side tools in DeepSeek Harness, Claude Code, and Codex without those calls being billed until October 31, 2026, at 16:00 UTC. The announcement does not make that waiver universal to every API use.

What are the rate limits?

The guide lists five queries per second for Search and five for Fetch at the account level, separately. Fetch also has a per-domain limit, so an application may encounter a 429 response even when its overall traffic appears modest.

References