DeepSeek made its name with cheap AI models. Now it’s undercutting search API rivals, too.

DeepSeek is pushing past cheap AI models and into search infrastructure. On Saturday, the company opened its in-house search engine to outside developers with two APIs, Search and Fetch, so apps and agents can pull live web results and page content directly. The price is simple: $0.75 for every 1,000 successful requests. The APIs can also be used without DeepSeek’s own models, which is the part that makes this move more than just another add-on.

That matters because LLMs stop at a cutoff date. Agents built for coding, research, or any other task that depends on current information need a way to check the web, and until now developers often stitched that together with third-party providers or DeepSeek’s own model-based search flow. The new setup cuts out an extra model request. Search returns ranked results with titles, URLs, snippets and summaries, while Fetch pulls webpage text in Markdown.

DeepSeek did not build this overnight. Its privacy policy in February 2025 still pointed to external search services such as Bing, and in March National Business Daily reported that it had integrated Bocha’s search API. Bloomberg later reported that DeepSeek was hiring engineers for a multilingual AI search engine, with image and audio support mentioned in job ads. Now the company says DeepSeek Search already handles hundreds of millions of requests a day and indexes hundreds of billions of webpages. It also says the system supports multilingual search and filters for spam, low-quality AI-generated material and ranking manipulation.

The pricing is the sharpest part of the story. DeepSeek is cheaper than many rivals, though not the absolute floor: Ceramic.ai starts at $0.25 per 1,000 queries, while Parallel, Perplexity, Exa, Brave and Tavily charge more on the figures cited by the company’s critics and competitors. At the same time, Google and Microsoft have been pulling back from standalone search APIs, which leaves room for these newer providers to sell direct access to results instead of keeping developers inside their own AI stacks.

There are limits, of course. Search and Fetch are capped at five requests per second per account, and DeepSeek only bills for requests that actually return results. The company is waiving charges for some server-side tool calls through DeepSeek Harness, Claude Code and Codex until October 31, but direct standalone calls are already billable. That makes this less like a grand search revolution and more like a very Chinese-style product move: undercut the market, ship the infrastructure, and let everyone else explain the margin.