Weekend Edition · Sat 17 Oct 2026
Beyond the Model: DeepSeek Opens Standalone Web Search and Fetch APIs for Agents
DeepSeek has launched Search and Fetch APIs that give AI agents live web results and page text for $0.75 per 1,000 successful requests, according to its documentation. That list price is well below the published rates of Brave, Exa and Perplexity, but the Journal has not tested retrieval quality, freshness or latency, so this article makes no performance claim.
Chanan Zevin - Chief Editor and Head of Desks
Published
What launched
DeepSeek has made two web-retrieval endpoints available to developers, according to its API documentation and a 17 October report by the Chinese technology site IT Home. The Search API takes a query and returns structured web results. The Fetch API retrieves a single web page and returns its content as Markdown. DeepSeek says both are designed for agents and retrieval-augmented generation applications. IT Home reported that DeepSeek officially announced the launch on the morning of 17 October, Beijing time.
How Search works
The documentation says Search is powered by DeepSeek's in-house search engine. A query can contain 1 to 128 characters, and a request can return between 1 and 50 results, with 10 as the default. Each result carries a title, a URL, a short snippet, a longer summary of 600 to 800 characters intended for a model, optional full page text and a publication date. It also carries a relevance score from 0 to 3 and an authority tier from 0 to 4. Developers can pass the user's original question to improve query understanding.
Operators for narrowing results
Search supports operators familiar from consumer search engines. The documentation lists restricting or excluding sites, exact-phrase matching, term exclusion, a language filter and date limits using "after:" and "before:". Date filters matter for agents, which often need to exclude stale pages when answering questions about current events. Whether DeepSeek's index applies those filters accurately is a testable question. The Journal has not yet tested it, and DeepSeek's documentation does not publish accuracy figures for them.
How Fetch works
Fetch takes one URL of up to 8,192 characters and returns the page title and its content in Markdown. The documentation says fetching "is subject to the target site's availability and access policies." It lists a 403 error for requests rejected by DeepSeek's security policy and a 404 error when no usable content can be extracted, including when a URL is blocked by site policy. That indicates some site restrictions are honoured. The documentation does not say whether the crawler follows robots.txt rules.
Billing per successful request
Both APIs are billed per successful request, not per token or per result. Search and Fetch each cost $0.75 per 1,000 successful requests, according to DeepSeek's English documentation. IT Home reported the price as 5 yuan per 1,000 requests. A request is billable only if it returns HTTP 200 with a non-empty result list or page content. Failed requests, empty searches and timeouts are not billed. The number of results requested does not change the price.
How the price compares
Chart 1 compares DeepSeek's list price with the published rates of three rivals for a basic search request. Brave lists its Search API at $5 per 1,000 requests. Perplexity lists its Search API at $5 per 1,000 requests, or $1 with its lower-latency Fast Search option. Exa lists $4 per 1,000 requests for its instant search and $7 for its fast and auto modes. The Journal read all of these figures from the vendors' own pricing pages on 17 October.
Why list prices mislead
The list prices are not like-for-like. Exa's base price covers up to 10 results and charges extra beyond that, while DeepSeek's covers up to 50. Perplexity bills a successful request even when it returns no results, while DeepSeek does not bill empty searches. Brave argues on its own pricing page that the real measure is the total cost of search and model tokens per correct answer. For builders, the meaningful comparison is cost per useful answer, and that can only come from testing.
Rate limits
The documentation sets an account-level limit of 5 queries per second for Search and a separate limit of 5 for Fetch. Fetch also has a per-domain limit. Requests above those limits receive an HTTP 429 error, and the limits apply regardless of how many API keys an account uses. By comparison, Brave's pricing page lists capacity of 50 queries per second for its Search plan. Low default limits matter for agent workloads that fan out many searches for each task. Enterprise terms may differ, but none are published.
Errors and retries
The documentation also sets out how failures should be handled, which matters for agents that run unattended. Errors follow the format used by OpenAI-compatible APIs. DeepSeek advises waiting and retrying after timeouts, rate-limit errors and server errors. It says an account with insufficient balance receives a 402 error and must be topped up first. Other client errors should not be retried without changing the request. Both APIs are read-only, the documentation says, so a retried call does not change anything on the target site.
Free inside rival coding tools
The documentation adds a notable condition. When developers use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, server-side tool calls already use Search and Fetch, and those calls are currently free of charge. Per-request billing for that route starts at 00:00 Beijing time on 1 November 2026, the documentation says. In the Journal's reading, the free window places DeepSeek's retrieval inside coding tools built by its rivals, not only inside its own products. Table 1 sets out the documented terms beside those of other providers.
| Provider | Search price per 1,000 | Page content | Billing and limits |
|---|---|---|---|
| DeepSeek | $0.75 | Fetch API: $0.75 per 1,000 | Empty results not billed; 5 QPS per API |
| Brave | $5 | Not separately priced on page | 50 QPS capacity listed |
| Exa | $4 instant; $7 fast/auto | Contents: $1 per 1,000 pages | Base price covers up to 10 results |
| Perplexity | $5; $1 Fast Search | fetch_url tool: $0.50 per 1,000 (Agent API) | Successful empty responses billed |
What has not been tested
A reproducible comparison of retrieval quality, freshness, latency and cost would be the most useful contribution a newspaper could make on this launch. The Journal has not run one. It did not hold API access to the new endpoints at the time of writing, and benchmark claims require a fixed query set, simultaneous runs and blind scoring. This article therefore compares documented features and published prices only. Any statement that DeepSeek's results are better or worse than a rival's would be unsupported at this stage.
A model entry that vanished
A separate report has caused confusion. IT Home said an X account called AI Tracker Bot spotted an entry labelled "deepseek-v4.1-flash-1017" in DeepSeek's official API model list, which was quickly removed. The account noted that the suffix matches DeepSeek's dated snapshot naming. It also said the public deepseek-flash endpoint still served V4.1 Flash, not a 1017 snapshot. DeepSeek had not explained the entry, IT Home reported. The Journal treats it as an observed listing, not as a launched model.
The September model release
The listing appeared a month after DeepSeek's last confirmed model release. IT Home reported that DeepSeek officially released V4.1 Flash on 10 September as the smallest model in its new architecture family, with native support for visual understanding. Developers call it by the name deepseek-flash. The previous V4 Flash and V4 Flash Vision Exp models were taken offline and temporarily routed to V4.1 Flash, according to the report. A dated snapshot would follow DeepSeek's past naming, but it has not been confirmed.
The harness release
DeepSeek also released version 0.2.1-rc1 of DeepSeek Harness, its agent software, on 17 October, according to IT Home. The release candidate adds plugin discovery from custom index sources, an experimental compatibility layer for Claude Code mods and automatic redaction of keys and tokens detected in sent messages, among other changes. IT Home said the Claude Code layer is intended to test whether that interface is broadly a subset of Harness plugins, not to provide full compatibility.
Owning the tools around the model
In the Journal's analysis, the launch fits a wider shift in which model developers sell the tools around the model as well as the model itself. Web search is among the most frequently called tools in agent workflows. Owning it gives a provider control over freshness, cost and the citations its agents produce. It also concentrates dependence on one company's index. The documentation reviewed by the Journal does not describe the scope of DeepSeek's crawling or the size of its index.
Questions for buyers
Developers considering the service face practical questions. Which languages and regions does the index cover well, and how fresh are its news results compared with rivals? How is the authority score assigned, and can it be audited? How long are queries retained, and where are they processed? The Search and Fetch documentation reviewed by the Journal does not address retention or data location. Buyers subject to data-protection rules will need to check DeepSeek's terms of service before sending user queries.
What to watch
Watch for independent benchmarks that test DeepSeek's retrieval quality and freshness against rivals on a fixed query set. Watch whether DeepSeek raises its five-queries-per-second default or publishes enterprise limits. Watch the start of billing on 1 November for searches routed through third-party coding tools, which will show how much usage the free window attracted. And watch whether DeepSeek explains the briefly listed 1017 model entry, which remains unconfirmed.
Sources
- DeepSeek API Docs, "Search & Fetch", read 17 Oct 2026 · api-docs.deepseek.com
- IT Home (IT之家), 17 Oct 2026, 14:36 Beijing time · ithome.com
- Vendor pricing pages read 17 Oct 2026 · brave.com exa.ai docs.perplexity.ai