DeepSeek Opens Its Search Engine to Agents: $0.75 per 1,000 Requests

DeepSeek has published documentation for two standalone APIs that give any agent or RAG pipeline access to its in-house search engine: Search, which returns reranked, structured web results for a query, and Fetch, which retrieves a single page as Markdown. Both cost $0.75 per 1,000 successful requests, are free until billing begins at midnight Beijing time on November 1, and are already wired in as server-side tools when DeepSeek runs inside DeepSeek Harness, Claude Code, or Codex. The story landed the day after the V4.1-Flash-1017 checkpoint, and together they sketch a lab building the full agent stack: a model that finishes tasks with shorter outputs, and a web layer priced for high-volume autonomous use.

The API shape: boring in the good way

Both tools are plain JSON REST endpoints on api.deepseek.com, authenticated with a standard DeepSeek key. Search is POST /v1/tools/search; queries run 1 to 128 characters and support real operators: site: and -site:, exact-phrase inclusion and exclusion, lang: filters, and after:/before: date bounds, with 1 to 50 results per call (default 10). Fetch is POST /v1/tools/fetch, takes only a URL (max 8,192 characters), and returns title plus Markdown content, subject to the target site’s own access policies. Rate limits are 5 QPS per account for each tool, separate from model concurrency, with 429s beyond that and an additional per-domain limit on Fetch.

What distinguishes the response schema is that it is built for model consumption rather than human search UI. Each result carries a rerank_score from 0 to 3 (the docs flag 1.5 and above as high quality), a UI-facing snippet of roughly 100 to 200 characters, a model-facing summary of 600 to 800 characters, an authority score from 0 to 4 with -1 for unknown, plus site_name, date in RFC 3339, and an optional full-page content field controlled by need_content. An agent can therefore filter on quality and authority scores before spending context tokens on any page, which is a meaningful token saving at scale.

The pricing model is the real signal

Billing is per successful request, not per token, per result, or per byte: a request only costs money if Search returns a non-empty result set or Fetch returns usable content on HTTP 200. Failures, empty searches, and timeouts are free, and count does not affect price, so requesting 50 results costs the same as 10. At face value, $0.75 per 1,000 requests is roughly an order of magnitude below the effective per-search cost of premium search-tool APIs, and it prices Fetch, the operation most likely to be blocked or rate-limited by publishers, at the same flat rate.

That number only makes sense because DeepSeek operates its own crawler and index rather than reselling someone else’s. It converts search from a metered luxury into a bulk commodity for agent workloads: a research agent that runs 200 searches and 500 fetches per task pays about half a dollar in tooling, which shifts the cost center entirely to model tokens. The competitive pressure lands on the incumbents that charge per-call premiums for search tools and on startups whose entire product is a search API wrapper.

What to watch

Three things determine whether this becomes infrastructure or a loss leader. First, index quality at the margins: the docs promise authority and rerank scoring, but agents in production will judge by whether DeepSeek’s index finds what Perplexity, Google, or Brave find, especially for fresh content and non-English queries. Second, publisher friction: Fetch “is subject to the target site’s availability and access policies,” and with websites increasingly closing their doors to AI agents, the actual yield per fetch attempt is the number that matters; a 403 costs nothing but wastes a model turn. Third, whether the free-until-November-1 window converts: server-side calls from Claude Code and Codex sessions are already running through these endpoints, so DeepSeek will have real usage data before the first invoice. If the quality holds, the November 1 switch-on is the moment agent search becomes a line item measured in dollars rather than budget meetings.