DeepSeek documents standalone Search and Fetch APIs

DeepSeek’s API documentation now describes two standalone endpoints for agents and retrieval. The English guide and the Chinese guide both name the same paths. Search is POST https://api.deepseek.com/v1/tools/search. Fetch is POST https://api.deepseek.com/v1/tools/fetch. Both use the same bearer API key as the rest of the DeepSeek API. The Search reference and the Fetch reference repeat those paths. A banner on the docs home page says the feature is available, powered by DeepSeek’s in-house search engine.

The change log, opened 17 October, has 10 September 2026 as its newest dated entry. That entry, for DeepSeek-V4.1-Flash, does not mention Search or Fetch. The guides give no launch date.

What a developer gets back

Search takes a query of 1 to 128 characters. count is optional, from 1 to 50, and defaults to 10. An optional ori_question, also up to 128 characters, is the user’s original question. An optional localization field is an IP address used to bias results. need_content defaults to false. When it is false, the content field is empty. Unknown fields in the body are ignored.

The query keeps its search syntax. The guide documents site: and -site:, a leading minus, quoted phrases, lang: with comma-separated codes, and after: and before: with a year or a full date.

Each result has a title, a URL, a short snippet, and a longer summary. The guide calls the snippet typically 100 to 200 characters and the summary typically 600 to 800. rerank_score runs from 0 to 3, and the guide treats 1.5 and above as high quality. authority runs from 0 to 4, with 3 and above called highly authoritative, and -1 when unknown. Date, site name, and site icon can be absent. A search with no matches returns an empty list.

Fetch takes one HTTP or HTTPS URL, at most 8,192 characters, with a host and no username or password. The response is a title, empty if none is available, and the page body as Markdown. The guide says the result depends on the target site’s availability and access policy.

The application makes the call and can read the scores before it puts text into a prompt. The guides do not publish prices for any other vendor’s search API.

Two price lines for the same unit

Both guides bill per successful request, not per token and not per result. A request is billed when it succeeds and produces a result: Search must return a non-empty list, and Fetch must return page text. Failed requests, empty searches, and timeouts are not billed. count does not change the price.

The numbers on the two language editions do not agree. The English pricing page and the English guide list Search at $0.75 per 1,000 successful requests and Fetch at the same $0.75. The Chinese pricing page and the Chinese guide list each at 5 yuan per 1,000 successful requests. Neither page explains the gap or says which figure applies to which account. Fees come off a topped-up balance or a granted balance, with the granted balance used first. DeepSeek says it can change prices.

The Chinese pages are where DeepSeek states a yuan price, Beijing time, and a Chinese public-holiday calendar. Model peak hours on that page follow Beijing weekday windows and skip Chinese public holidays. Those hours apply to model tokens. Search and Fetch stay on the per-request price.

Who is free, and for how long

Both the English and the Chinese pages say the same thing about a waiver. When the DeepSeek API is used through DeepSeek Harness, Claude Code, or Codex, server-side tool calls already hit Search and Fetch. Those calls are free of charge now. Per-request billing starts at 00:00 Beijing time on 1 November 2026. The sentence is about using the DeepSeek API inside those tools. It does not say that Claude Code or Codex, pointed at some other provider, routes its web search to DeepSeek.

Pandaily, opened the same morning, reports 5 yuan per 1,000 successful calls, which matches the Chinese pages, and a cap of 5 queries per second, which matches both guides. It also says Search fees inside Claude Code and Codex are waived for the first two weeks after launch, with the billing start date to be announced later. The docs opened on 17 October name 1 November 2026 at midnight Beijing time. This account follows the docs on the date.

The rate cap

Search and Fetch are limited separately, at 5 requests per second each, counted per account rather than per API key. Traffic on one does not spend the other’s quota. The cap is separate from model concurrency. Fetch has a further per-domain cap. Requests over either cap are refused. The guide says to wait and retry after a rate-limit response, a timeout, or a server error, and to top up after an insufficient-balance response.

A developer has a documented pair of endpoints, a 5-per-second account cap, and a written free window for Harness, Claude Code, and Codex tool calls through 31 October. The price to budget is $0.75 or 5 yuan per 1,000 successful calls, and the docs do not say which figure an account is billed.