DeepSeek today opened its Search and Fetch APIs to all developers, charging $0.75 per 1,000 calls for each endpoint and naming October 31 as the day server-side tool calls from Claude Code, Codex and its own DeepSeek Harness stop being free.

In Short

DeepSeek, the Chinese AI company, now sells two tools that let software look things up on the web: one finds pages for a question, the other reads a single page and hands back its text. Anyone building AI assistants or agents can pay DeepSeek about 75 cents for every 1,000 lookups, which is far cheaper than what Microsoft charges for a comparable service. If you use DeepSeek's models inside coding tools such as Claude Code or Codex, the web lookups those tools make on your behalf are free today, but they start costing money at 16:00 UTC on October 31, 2026, which DeepSeek's own documentation describes as midnight on November 1 in Beijing.

What DeepSeek published

The announcement arrived as the third post in a four-part thread from the company's official X account, @deepseek_ai, timestamped 10:35 AM on October 17, 2026. "Search & Fetch APIs are now open to everyone!" the post reads. It describes the service as "Powered by our in-house search engine, serving hundreds of millions of requests daily worldwide," gives a price of "$0.75 per 1,000 calls" for each endpoint, and points to the documentation at api-docs.deepseek.com/guides/search. The post had drawn 291,200 views by the time it was captured.

The thread itself opened with something else entirely. Its first post introduced DeepSeek-V4.1-Flash-1017, a model the company said delivered "Stronger benchmark performance, faster task completion, and shorter outputs than the previous version," and which it described as live on its website, app and API. The second post said the model "uses reinforcement learning with self-feedback to improve end-to-end project development." The search announcement came third, which means it was folded into a model launch rather than given a thread of its own. The fourth post was not part of the material reviewed for this article.

Three pages from the DeepSeek API documentation accompany the post: a guide titled "Search & Fetch," and two API reference pages, one for each endpoint. None of the three carries a publication date. All three display the same site-wide banner: "Search / Fetch API is now available, powered by our in-house search engine and supporting web search and fetch." The guide sits in the documentation's navigation with a "NEW" label, between the Files API and Context Caching.

The documentation frames the purpose plainly. "The Search API and the Fetch API give agents and RAG applications access to live web content," according to the guide. Search, in DeepSeek's words, "finds web pages for a query and returns structured results," while Fetch "retrieves a single web page and returns its content as Markdown."

Two endpoints, one key

Both services are described as "raw JSON APIs" that "use the same bearer authentication as the DeepSeek API." There is no separate product key. A developer already paying for DeepSeek's models sends an API key in an Authorization header, a content type of application/json, and posts to one of two addresses:

The /v1/tools/ path is worth noticing. DeepSeek places these endpoints alongside its chat and completion APIs rather than as a separate search product, and the guide's navigation also lists sections titled "Using the Responses API" and "Using the Anthropic API." The company is positioning its model API as a drop-in backend for tools built around other providers' formats, and search is part of that bundle.

The Search request

The Search endpoint accepts five fields, according to both the guide and the API reference.

query is the only required field. It must contain 1 to 128 Unicode characters and "cannot be all whitespace." count sets the number of results, from 1 to 50, with a default of 10. ori_question carries "The user's original question that triggered the search," also limited to 1 to 128 characters; the reference page says providing it "can improve query understanding" and that null "is treated as not provided." localization is described as "An optional IP address used to localize search results," defaulting to an empty string. need_content is a boolean that controls whether full page content is returned; it defaults to false, and when false the content field "is always an empty string."

"Unknown fields in the request body are ignored," the guide adds.

Two of those fields carry implications beyond relevance. The ori_question parameter invites an application to forward the end user's own words to DeepSeek alongside the reformulated query. The localization parameter asks for an IP address. Neither page says how DeepSeek stores or uses either value. That gap matters for a company whose consumer app has faced regulatory action in Europe: Italy's Garante banned DeepSeek on January 30, 2025, and the Berlin Data Protection Commissioner declared its operations unlawful on June 27, 2025, with both cases tied in part to the transfer of personal data to China.

Query operators

According to the guide, "The query is forwarded with its search syntax intact," and operators can be combined in a single query. Eight are documented:

  • site: restricts results to a site; multiple sites are joined with OR, as in DeepSeek site:github.com OR site:huggingface.co
  • -site: excludes a site
  • a leading hyphen excludes pages containing a term
  • -"..." excludes pages containing an exact phrase
  • "..." matches an exact phrase
  • lang: restricts results to language codes, separated by commas, such as lang:zh,en
  • after: restricts results to pages published after a year or a full date
  • before: restricts results to pages published before a year or a full date

The guide's combined example reads: DeepSeek site:github.com OR site:huggingface.co lang:en after:2025. The Search reference page describes the same feature more narrowly, saying the query "Supports operators such as site:, -site:, and quoted phrases" and sending readers to the guide for the full list. The lang:, after: and before: operators appear only in the guide.

What a search returns

A successful call returns HTTP 200 and an array called results. "A normal search with no matches returns an empty array," according to the guide. Each result carries nine fields.

rerank_score is "A relevance score from 0 to 3 that combines signals such as relevance, authority, and freshness," with "scores of 1.5 or above considered high-quality results." The guide types it as a number; the reference page types it as a float and lists only an upper bound of 3. title and url are self-explanatory. snippet is "A short summary related to the query, typically 100 to 200 characters, intended for display in a UI." summary is "A long summary related to the query, typically 600 to 800 characters, intended to be read by a model." content holds page content when need_content is true and the content is available. site_name and site_icon may be null. date is "The publication time in UTC, formatted as RFC 3339, or null when unknown." authority is "The authority tier of the page, 0 to 4 (-1 means unknown)," with "values of 3 or above indicating highly authoritative sources."

The split between snippet and summary is a design choice that says something about the intended buyer. One field is written for humans looking at a results list. The other, several times longer, is written for a language model to read before answering. DeepSeek is returning pre-digested text so the calling model does not need to open every page, a pattern that reduces the number of downstream fetches but also means the publisher's page may never be loaded by the agent at all.

The authority score is equally notable. DeepSeek exposes a five-tier ranking of how authoritative it considers each site, but neither page explains how a page earns a tier. For publishers, that number is now a visible judgment by a search engine about their standing, delivered to every developer who calls the API, with no documented method behind it.

The Fetch request

Fetch is simpler. It takes one required field, url, which according to the guide must be "at most 8192 characters, an HTTP or HTTPS URL with a host, and must not contain a username or password." The API reference adds that the value must be non-empty. The response contains two fields: title, which is "empty when the title is unavailable," and content, "The fetched page content as Markdown text."

"Fetching is subject to the target site's availability and access policies," the guide states. It does not name the user agent the fetcher presents, publish IP ranges, or say whether robots.txt is consulted. The error table offers indirect evidence: HTTP 404 is returned when Fetch "could not retrieve usable page content," and the listed causes include "a URL blocked by site policy." A separate HTTP 403 applies when "The Fetch request was rejected by the security policy," which appears to refer to DeepSeek's own policy rather than the target site's.

That ambiguity places DeepSeek in a familiar position. Site owners deciding how to treat AI traffic generally rely on declared crawler names and robots.txt rules. Without a documented identity, a Fetch request is indistinguishable in the documentation from any other automated visitor, and publishers have no published way to allow or block it specifically.

Errors and retries

Errors follow "the OpenAI-compatible error format," the guide says, adding that "The param field is always null." The documented example is a request with an out-of-range count, which returns the message "'count' must be between 1 and 50" with a type and code of invalid_request_error.

The status code table covers ten cases: 400 for invalid headers or bodies, 401 for a missing or invalid key, 402 for insufficient account balance, 403 for security policy rejections, 404 for unusable Fetch content, 408 when reading the request body times out, 413 for an oversized body, 429 when "The Search or Fetch QPS limit was exceeded," 500 for internal server errors, and 504 when "A downstream request timed out."

"Search and Fetch are read-only," the guide notes. It recommends waiting and retrying on 408, 429 and 5xx responses, topping up an account before retrying a 402, and changing the request before retrying any other 4xx error.

Rate limits

Each endpoint is capped at five queries per second per account. "For each account, Search and Fetch are limited independently, and traffic to one does not consume the other's quota," according to the guide. The limits apply "at the account level, regardless of which API key is used," and they are "separate from the model concurrency limits."

Fetch carries a second constraint. "In addition to the account-level limit, Fetch is also subject to a per-domain QPS limit, and requests above that limit also receive HTTP 429." The guide does not say what that per-domain limit is.

Five requests per second amounts to 432,000 calls a day at sustained maximum, a ceiling suited to individual applications and coding agents rather than bulk crawling. Set against the X post's claim that the underlying engine serves "hundreds of millions of requests daily worldwide," the per-account cap signals that DeepSeek is metering access to its index rather than opening it for wholesale extraction. The "hundreds of millions" figure appears only in the X post; the documentation pages do not repeat it, and DeepSeek offers no breakdown of where that volume comes from.

How billing works

The guide is explicit that pricing is per request, "not by tokens or by the number of returned results." Search and Fetch each cost $0.75 per 1,000 successful requests.

A request counts as billable only when it "returns HTTP 200 and produces a result." For Search, that means a non-empty results array; for Fetch, returned page content. "Failed requests, empty search results, and timeouts are not billed," the guide states, and "count only controls how many results are returned and does not affect the per-request price."

The flat price creates an incentive that is easy to miss. A call requesting 50 results costs the same as a call requesting one. An agent configured to pull the maximum on every query pays nothing extra, and the snippet and summary fields arrive for every one of those 50 results.

Where the X post and the documentation differ

The X post and the guide describe the same pricing in slightly different terms, and the differences are worth recording.

The X post says "$0.75 per 1,000 calls." The guide says "$0.75" per "1,000 successful requests" and then defines success narrowly. A reader working from the X post alone would assume every call is billed; the documentation says empty and failed calls are not.

The two sources also state the end of the free period differently. The X post says: "Search & Fetch calls via server-side tools in DeepSeek Harness, Claude Code, and Codex are currently free for DeepSeek API users. Billing starts Oct 31, 2026, at 16:00 UTC." The guide says those calls "are currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026." Beijing time runs eight hours ahead of UTC, so both describe the same moment. The calendar date differs depending on which document is read, though, and a team in Europe or the Americas reading only the guide could plan for November 1 and be billed for part of October 31.

The X post uses the word "open to everyone," while the documentation banner says the service is "now available." Neither source mentions a free tier, trial credit or waitlist for direct API calls; the only free use described in either document is the server-side tool path through the three named clients, and that ends on October 31.

The Claude Code and Codex detail

The most consequential line in the documentation may be a note at the end of the billing section. "When you use the DeepSeek API through DeepSeek Harness, Claude Code, or Codex, server-side tool calls already hit the Search and Fetch APIs," according to the guide.

In other words, developers who have pointed Anthropic's Claude Code or OpenAI's Codex at DeepSeek's models are already using DeepSeek's search engine whenever those tools run a web search or fetch a page on the server side. The traffic has been flowing before the public announcement. What changes on October 31 is that it starts costing money.

This explains the "Using the Anthropic API" and "Using the Responses API" sections in the documentation's navigation. DeepSeek exposes endpoints compatible with the formats used by Anthropic and OpenAI, so their coding agents can run on DeepSeek models. When those agents invoke a built-in web search tool, the request lands on DeepSeek's infrastructure rather than on the search provider the tool's maker would normally use. The first post in the thread also told users to "Upgrade to DeepSeek Harness v0.2.1-rc.1 or later for the best experience with the new model," linking the model release, the harness and the search service into one package.

The billing change is narrow. It affects DeepSeek API customers who route these tools through DeepSeek, not Claude Code or Codex users in general.

DeepSeek is entering a market that has been repriced several times in the past 18 months, and the comparison points are instructive.

Microsoft retired the Bing Search and Bing Custom Search APIs on August 11, 2025, steering developers to Grounding with Bing Search at $35 per 1,000 transactions, an increase of 40% to 483% over the old tiers depending on plan. That replacement did not return raw results; a model processed them and returned a cited answer. Microsoft later added a second product, Web IQ, described on June 2, 2026, a grounding API, which returns passage-level evidence rather than full documents.

Brave sells its search API for $5 per 1,000 requests on its Search plan, with $5 of monthly free credit conditional on attribution. Cloudflare's Web Search API, announced October 2, 2026, routes agent queries to Ceramic.ai at $0.25, Linkup at $5.00 and Exa at $7.00 per 1,000 requests, with queries of up to 1,024 characters and up to 10 results per request. DeepSeek's limits run the other way: a much shorter query cap of 128 characters, but up to 50 results.

At $0.75, DeepSeek prices its search below Brave and every Cloudflare partner except Ceramic.ai, and at roughly one forty-seventh of Microsoft's grounding rate. The two are not equivalent products, since Microsoft's offering includes model processing. Still, for a developer who only needs ranked links and summaries to feed into retrieval-augmented generation, the cost difference is large.

The Fetch side competes with a different set of tools. SerpApi's Extract API, covered on October 13, 2026, also returns Markdown and can crawl up to 100 pages per request, but it does not execute JavaScript and does not document whether it honours robots.txt. DeepSeek's Fetch retrieves one page per call and is equally silent on JavaScript rendering and crawler identity.

The coincidence of dates is also hard to miss. Cloudflare's AI Search product begins billing on November 1, 2026, with semantic queries priced at $0.75 per 1,000. The same price point for a query and almost the same start date land in the same fortnight from two unrelated companies.

Why this matters for publishers and marketers

Every Search response hands a model a 600 to 800 character summary of a page, written by DeepSeek, before the model decides whether to open anything. For publishers, that is another layer of the dynamic PPC Land has tracked through Cloudflare's crawl-to-referral data, which showed ratios as high as nearly 50,000 crawls per referred reader. An agent satisfied by the summary never sends a visitor.

The Fetch endpoint creates a second issue. Agent traffic, the category Cloudflare describes as fetches made for a specific user during a live interaction, is the hardest kind for site owners to manage, and DeepSeek has not published how its fetcher identifies itself. When Cloudflare moved toward blocking Training and Agent crawlers by default on ad-carrying pages for new domains from September 15, 2026, it did so on the basis of crawler classification. A fetcher without a documented name sits outside that system unless the network operator identifies it some other way.

There is also a measurement angle. DeepSeek's share of AI assistant usage remains small in Western markets: Comscore data covered on September 26, 2026, put DeepSeek at 2% of US prompt volume in June 2026, up from 1% in January. The Search and Fetch APIs may route a much larger volume of DeepSeek-powered lookups through third-party tools than its own chatbot ever shows. Search engine visibility work, much of it now aimed at being cited by AI systems, has largely focused on Google, Microsoft and OpenAI. A cheap search API from a Chinese provider, with its own authority tiers and its own index, adds a source whose ranking signals are undocumented and whose summaries publishers cannot see.

For advertisers, the effect is indirect but real. Search results delivered through an API carry no ads, and summaries consumed by a model carry no ad impressions. The more agent queries shift to low-cost endpoints like this one, the less of that research activity happens on pages where ads are served.

Timeline

Summary

Who: DeepSeek, the Chinese AI developer, through its official X account and its API documentation. Developers building agents and retrieval applications, DeepSeek API customers using Claude Code, Codex or DeepSeek Harness, and publishers whose pages are searched and fetched.

What: DeepSeek opened two endpoints, Search and Fetch, to all API users. Search returns up to 50 ranked results per query with scores, authority tiers, short snippets and longer model-oriented summaries; Fetch returns a single page as Markdown. Each costs $0.75 per 1,000 successful requests, with a five-requests-per-second limit per account per endpoint and an undisclosed per-domain limit on Fetch. Server-side tool calls from Claude Code, Codex and DeepSeek Harness, which already use these endpoints, are free until billing begins.

When: Announced today, October 17, 2026, at 10:35 AM in a post that was third in a thread alongside the DeepSeek-V4.1-Flash-1017 model release. Billing for server-side tool calls starts at 16:00 UTC on October 31, 2026, which the documentation states as 00:00 Beijing time on November 1, 2026. The documentation pages are undated.

Where: The endpoints sit at api.deepseek.com/v1/tools/search and api.deepseek.com/v1/tools/fetch, and are described in DeepSeek's API documentation. The X post says the underlying search engine serves requests worldwide.

Why: DeepSeek is extending its model API into the infrastructure that AI agents use to reach the live web, at a price below Brave, most Cloudflare partners and Microsoft's grounding service. The move matters to publishers and marketers because summaries written for models reduce the need for agents to visit pages, the fetcher's identity is undocumented, and the authority scores DeepSeek assigns to sites carry no published method.