DeepSeek is now renting out its search engine. The Chinese AI lab has opened the in-house web search powering its products to outside developers through two new endpoints — a Search API that returns ranked web results and a Fetch API that pulls a single page back as clean Markdown — each priced at $0.75 per 1,000 successful requests. The launch, announced on October 17 alongside the company's V4.1-Flash-1017 model update, is DeepSeek's latest move in a pricing war that keeps pushing the cost of building AI agents toward zero.
DeepSeek says the underlying engine indexes hundreds of billions of web pages and, according to its changelog, already handles hundreds of millions of search requests from around the world every day. Search and Fetch are aimed squarely at agent builders: ranked web results for grounding answers, and Markdown extraction for retrieval-augmented generation pipelines and research automation. Failed requests, empty results and timeouts are not billed, and the number of results requested does not change the price.
A fraction of rival list prices
At $0.75 per 1,000 calls, DeepSeek sits well below the going rate. Brave and Perplexity list their search APIs at $5 per 1,000 requests, Kagi charges $12, Exa starts at $4, Tavily's pay-as-you-go works out to roughly $8, and Google charges $14 per 1,000 Grounding-with-Google-Search requests on Gemini 3 and newer models. The products are not identical — Brave bundles extra LLM context, Google's grounding runs inside a model call — so list price is not the whole story. But as a pure per-query search cost, DeepSeek's number is the cheapest from any major provider.
It fits a pattern. DeepSeek has repeatedly paired new releases with aggressive price cards to pull developers onto its platform, and this time the target is the tooling layer agents run on, not the models themselves. Both OpenAI and Anthropic have cut API prices this year; DeepSeek's bet is that search — the toll booth every browsing agent must pass through — is the next place margins collapse.
Free until November 1, then the meter starts
There is a deadline attached. Server-side calls made through the DeepSeek Harness, Claude Code or Codex are currently free, but per-request billing begins at 00:00 Beijing time on November 1, 2026 (16:00 UTC on October 31). After that, every search plus every page fetch counts as a billable request. DeepSeek caps each account at 5 requests per second for Search and 5 for Fetch, applies a per-domain limit on Fetch, and uses the same API key and base URL developers already use for its models — wiring live web results into an agent is mostly a configuration change.
Developers have noticed. The launch thread on Hacker News had crossed 543 points by October 19, with commenters probing how Fetch handles anti-bot protection and whether DeepSeek will pay the creators of fetched pages. Some gaps remain in the documentation — the changelog gives no index size, no supported-language list and no detail on spam and AI-content filtering, and the per-domain fetch limit is unquantified — but for agent builders the math is simple. Cheap search is the cheapest way to make a smart agent cheaper.
