News · 2026-10-18

DeepSeek opens Search and Fetch at $0.75 per thousand successful calls

DeepSeek opened its in-house Search and Fetch programming interfaces on October 17, charging $0.75 per 1,000 successful requests for each endpoint. Search returns structured web results, while Fetch retrieves a specified page as readable text. The launch gives agent developers a directly priced way to obtain current information, with a billing rule that excludes failed requests, empty searches, and timeouts.

Key facts

• Each endpoint costs $0.75 per 1,000 successful requests, independently of model-token charges.
• DeepSeek’s official changelog dates the opening to October 17, 2026.
• Search requires a successful response with non-empty results to incur a charge; Fetch requires returned page content.
• The primary source is DeepSeek’s Search and Fetch guide, alongside its dated changelog.

A missing piece of the agent stack becomes a product

A language model’s stored knowledge cannot tell it what changed on a website this morning. An agent needs a separate mechanism to search for relevant pages and then read them. DeepSeek is now offering those two operations as services developers can call directly using their existing DeepSeek access key.

The company’s changelog describes the move as “Opening up the in-house search API.” That wording matters because this is a tool launch alongside the model service, rather than a claim that the language model itself has acquired permanent knowledge of the live web.

Search accepts a query and returns structured results. It can optionally include content. Fetch starts with a known page address and returns its content in Markdown, a text format that preserves elements such as headings and links. The difference resembles asking a librarian which books address a question versus requesting the relevant pages from a book already selected.

Both operations can support retrieval-augmented generation: the software obtains external material and gives it to the model before asking for an answer. The quality of that material still determines what the model can responsibly say.

The price is simple; the workflow still needs counting

DeepSeek defines successful calls precisely. Search becomes billable when it returns a successful response containing at least one result. Fetch becomes billable when a successful response contains page content. Errors, empty search results, and timeouts are not charged. Asking Search for a different number of results does not change that request’s price.

For an illustrative workload, 1,000 successful searches plus 1,000 successful page fetches would cost $1.50 in these tool fees. That is arithmetic from the published request price, not a quoted package or a measurement of a real application. Model inference, additional searches, and any other services remain separate costs. An agent that repeatedly reformulates queries can make many calls while answering one user question.

The guide specifies separate account-level ceilings of five requests per second for Search and Fetch, with an additional per-domain limit for Fetch. Those constraints matter for parallel agents: a low unit price does not imply unlimited simultaneous retrieval.

DeepSeek also documents a transition for server-side tool calls through DeepSeek Harness, Claude Code, and Codex. Those calls are currently free, with billing scheduled to start at midnight Beijing time on November 1, equivalent to 16:00 UTC on October 31. Developers should distinguish that temporary treatment from the direct endpoint price. The company’s pricing documentation and guide provide the commercial terms.

Retrieval quality is the real competitive test

DeepSeek says its search engine is built in-house and filters search spam and material designed to manipulate answers from generative systems. The documentation does not provide an independent evaluation proving those quality claims. A cheap search that misses the authoritative page can cost an agent more reasoning and retries, or lead it to produce a worse answer.

The Hacker News discussion includes early reports of mixed results, including an unsuccessful search for a Microsoft product landing page and a query that placed Debian repository pages above a small project’s homepage. These are individual tests, not a representative comparison. They nevertheless identify the right question: whether the service consistently finds the source an application actually needs.

The launch coincided with DeepSeek’s Flash-1017 update, which Ground Truth covered as a separate model refresh. Bundling the two into one capability claim would obscure the useful distinction between a model and the tools it can call.

For builders, the immediate opportunity is straightforward integration and predictable successful-call pricing. The honest caveat is equally practical: test relevance, page extraction, and total calls per completed task before treating the headline request price as the cost of a reliable research agent.

Primary source, verified: read the paper →

Ground Truth. (2026, October 18). DeepSeek opens Search and Fetch at $0.75 per thousand successful calls. Ground Truth. https://groundtruth.day/news/deepseek-search-fetch-success-based-pricing.html

@misc{groundtruth:deepseek-search-fetch-success-based-pricing,
  title  = {DeepSeek opens Search and Fetch at $0.75 per thousand successful calls},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {oct},
  url    = {https://groundtruth.day/news/deepseek-search-fetch-success-based-pricing.html}
}

Topics: tools · search · agents · deepseek · retrieval