DeepSeek's Search and Fetch APIs officially launched on October 17, 2026, with DeepSeek opening the search engine behind its own app to all developers through its official API documentation. Previously, the only way to use DeepSeek's search programmatically was the web search built into its own products; now search and page retrieval are split into two standalone endpoints that any agent or RAG application can call, with no binding to DeepSeek's models.
Search finds, Fetch retrieves
Search takes a query and returns structured results: title, link, snippet, a longer model-oriented summary, optional full text, publication date, plus two fields worth noting — a reranking score and a source authority rating. Queries are capped at 128 characters, and one call returns between 1 and 50 results, 10 by default. Fetch takes a single web page and returns its cleaned title and body as Markdown. Both are raw JSON APIs sharing the api.deepseek.com domain and the existing Bearer key, so for teams already holding a DeepSeek API key the integration cost is close to zero.
Billing counts only successes; failures and empty results are free
Both endpoints cost the same: 5 yuan per thousand successful requests. The billing rules are unusually developer-friendly — an empty Search result is not charged, a Fetch that returns no body text is not charged, failures and timeouts are not charged, and the number of results returned does not affect the price. Rate limits are counted separately per account for each endpoint, at 5 requests per second each; Fetch adds a per-domain limit on top, and exceeding it returns HTTP 429. For comparison, comparable overseas search APIs commonly charge 5 to 10 US dollars per thousand calls, roughly an order of magnitude more — a difference that matters quickly for high-volume agent applications.
Two free weeks at launch: the fight is over the agent retrieval layer
The documentation adds a time-limited arrangement: until November 1, 2026, Search calls triggered through Claude Code's Web Search tool or Codex's web_search function are free. The intent is plain — get the major coding agents running their retrieval traffic through it first, and talk about charging later. Search is becoming a standard agent accessory, and a model vendor selling search itself is selling more than data: it is selling the judgment layer of ranking and authority filtering, with an index that DeepSeek says filters out SEO spam and low-quality AI content. Two caveats deserve stating: the authority-rating criteria are not public, and no third party has evaluated the index's coverage, so real quality will only be known once developers run real queries against it. For teams already on the DeepSeek API, this is the most convenient add-on available today; for teams built on other models, it is also a standalone low-cost retrieval option.
