The Pulse
DeepSeek Adds Standalone Search and Fetch APIs for Agents
DeepSeek’s standalone Search and Fetch endpoints let developers call web discovery and page retrieval separately, without requiring a model response to initiate a search. The APIs cost $0.75 per 1,000 successful requests, while qualifying s
AI.info Team · Written with AI ·
DeepSeek already offers server-side web search through its Responses API, but that route is limited to the deepseek-v4-pro model. Its newly documented standalone Search and Fetch endpoints split web discovery and page retrieval into separate calls, giving developers a more direct way to add those steps to agents and retrieval-augmented generation systems.
The distinction matters for teams building their own workflows: the new APIs do not require a model response to initiate a search. DeepSeek’s documentation describes Search as using the company’s in-house search engine, while Fetch takes a URL and returns the page content as Markdown. Both use raw JSON requests and the same bearer-token authentication as the company’s other APIs.
Search returns ranked pages; Fetch reads one URL
Search accepts a required query and can return between one and 50 results; the default is 10. Developers can request page content in the search response, pass an optional original question to improve query understanding, or provide an IP address for localized results. The documented query syntax supports operators such as site:, exact-phrase matching, language filters and date limits.
Each result can include a title, URL, snippet, longer summary, publication date, authority tier and rerank score. The guide describes the rerank score as a 0-to-3 measure combining relevance, authority and freshness, and says scores of 1.5 or above count as high quality. When developers need the full text of a particular page, Fetch accepts its URL and returns a title plus Markdown content; the URL must be HTTP or HTTPS and no longer than 8,192 characters.
Successful calls cost $0.75 per 1,000
DeepSeek’s English-language guide prices Search and Fetch separately at $0.75 per 1,000 successful requests. A Search call counts only if it returns HTTP 200 with at least one result; Fetch counts only when it returns page content. Empty searches, failed requests and timeouts are not billed, and asking Search for more results does not change its per-request price.
The Chinese-language version lists both services at ¥5 per 1,000 successful requests. The guides do not explain whether the two published prices reflect different billing arrangements or simply present the rate in each language’s currency. Developers should check the pricing displayed for their account before estimating costs.
Five requests per second, with Fetch limits by domain
Search and Fetch each have an account-level limit of five requests per second, independent of one another and separate from model concurrency limits. Fetch also has a per-domain limit, so requests concentrated on one site may receive HTTP 429 errors even when the account-wide rate is below its cap. DeepSeek classifies the APIs as read-only and advises retrying timeouts, rate-limit responses and server errors, but not other client errors without changing the request. For HTTP 402, the guide says to top up the account before retrying.
Fetch may also fail when a site is unavailable, blocks access, returns an unsupported content type or has no extractable Markdown. That makes the endpoint a page-reading service subject to the destination site’s availability and access policies, not a guarantee that any URL can be retrieved.
Standalone tools sit beside model-led search
The Responses API guide documents a separate, model-led web_search tool for deepseek-v4-pro, where DeepSeek runs the search, adds results to the model’s context and continues the response. Standalone Search and Fetch instead expose discovery and page retrieval as individually callable APIs, so developers can decide when to search, which result to fetch and how to pass the returned text into their own system.
DeepSeek says server-side tool calls to both Search and Fetch through DeepSeek Harness, Claude Code or Codex are currently free. Per-request billing for those calls starts at 00:00 Beijing time on November 1, 2026. The documented endpoints are POST /v1/tools/search and POST /v1/tools/fetch.