DeepSeek opens search engine to developers, undercutting AI search API rivals

New Search and Fetch endpoints deliver live web results for $0.75 per 1,000 requests and work with any AI model

By LineZotpaper

Published20 October 2026

Read Time3 min

Chinese AI company DeepSeek has opened its in-house search engine to outside developers, launching two APIs that retrieve live web results and webpage content for $0.75 per 1,000 successful requests. The move puts DeepSeek in direct competition with search API providers such as Perplexity, Exa and Brave as developers build AI agents that depend on current information.

DeepSeek, best known for producing low-cost frontier AI models, announced on Saturday that its Search and Fetch APIs are now available to developers. Search returns ranked web results including titles, URLs, snippets and summaries, while Fetch retrieves the contents of individual webpages in Markdown format. "Neither requires a separate model request, meaning developers can integrate DeepSeek's search capabilities into applications powered by competing AI models, too," the company said. The launch was accompanied by DeepSeek-V4.1-Flash-1017, a minor update to the V4.1 Flash model released in September.

Large language models are trained on data collected up to a fixed point in time, which limits their answers on recent events, updated documentation and other dynamic information. Live web search makes that information available to models, and it is particularly important for AI agents, such as coding agents that need to check the latest software documentation or newly disclosed vulnerabilities.

Previously, developers using DeepSeek's models could connect third-party search services or use DeepSeek's existing search functionality through its model API, which required an additional model request per search and added latency and token costs. The new endpoints bypass that step.

DeepSeek has relied on outside providers for much of the past two years. Chinese publication National Business Daily reported in March 2025 that DeepSeek had integrated the search API of Bocha, a Hangzhou-based AI search startup, and DeepSeek's February 2025 privacy policy acknowledged third-party services such as Microsoft's Bing. Bloomberg reported in January that DeepSeek was recruiting engineers to build a multilingual AI search engine, with plans extending to image and audio search.

According to an October 17 changelog, the resulting engine, dubbed DeepSeek Search, already handles hundreds of millions of requests daily and maintains an index containing hundreds of billions of webpages. It supports multilingual searches and uses filtering mechanisms designed to identify search spam, low-quality AI-generated material and attempts to manipulate search rankings. DeepSeek has disclosed little about the engine's underlying architecture or whether it has entirely replaced its external search suppliers.

The launch comes as the market for standalone search APIs grows. Brave operates an independent index of more than 40 billion webpages, Exa has built its own crawling and indexing infrastructure, Perplexity has invested in proprietary search systems, and Parallel says it has built its stack from the ground up. Cloudflare entered the market this month with a Web Search API through its AI Gateway. Meanwhile, Microsoft retired its Bing Search APIs in August 2025, and Google has closed its Custom Search JSON API to new customers, with existing users required to migrate by January 1, 2027.

On price, DeepSeek's $0.75 per 1,000 searches compares with $1 per 1,000 for Parallel's Turbo and Fast services and Perplexity's Fast Search API, $4 for Exa Instant Search, $5 for Brave, and $8 for Tavily. Ceramic.ai charges $0.25 per 1,000 queries on its pay-as-you-go plan. The services differ in retrieval capabilities, result limits and performance, so the prices are not entirely comparable.

Analysis

Why This Matters

DeepSeek's low-cost entry could put pricing pressure on the AI search API market, following its pattern of disrupting frontier AI labs with cheaper models.
AI developers gain a low-priced, direct way to give agents live web access without depending on Google or Microsoft's AI platforms.
Web search is becoming strategic infrastructure for AI agents, making DeepSeek's move part of a broader competitive shift in the industry.

Background

Large language models cannot answer questions about events after their training data was collected, which is why agent builders connect them to live web search. Operating a search engine requires crawling the web, maintaining an index, ranking results and continuously updating that information, work that for years was dominated by Google and Microsoft at web scale. A newer generation of companies, including Brave, Exa, Perplexity, Parallel and Cloudflare, has invested in its own search infrastructure with AI applications in mind. DeepSeek built its engine after relying on third-party search services, and has now opened it to developers at the same time that Google and Microsoft have pulled back from standalone search APIs.

Key Perspectives

Developers: Gain a cheap, direct search API that works with any AI model, with no separate model request and lower latency than routing through DeepSeek's model API. Incumbent search API providers: Face a price competitor backed by an established AI company, though they can argue that retrieval quality, result features and performance justify higher prices. Critics/Skeptics: DeepSeek has revealed little about how DeepSeek Search is built or maintained, and has not made clear whether it has completely replaced external search suppliers, leaving questions about reliability and index quality.

What to Watch

How DeepSeek Search evolves, including the image and audio capabilities it reportedly planned.
Whether competing search API providers respond with price cuts or differentiate on features.
The January 1, 2027 deadline for Google Custom Search JSON API users to migrate, which could push more developers toward standalone providers such as DeepSeek.

Sources

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

Checked

12 claims in this story checked against 1 source: 11 supported.

Show each claim

SupportedDeepSeek opens search engine to developers, undercutting AI search API rivalsThe New Stack: “DeepSeek made its name with cheap AI models. Now it’s undercutting search API rivals, too.”
SupportedNew Search and Fetch endpoints deliver live web results for $0.75 per 1,000 requests and work with any AI modelThe New Stack: “Both services cost $0.75 per 1,000 successful requests”
Supported"Neither requires a separate model request, meaning developers can integrate DeepSeek's search capabilities into applications powered by competing AI models, too," the company said.The New Stack: ““Neither requires a separate model request, meaning developers can integrate DeepSeek’s search capabilities into applications powered by competing AI models, too.””
Partly supportedAccording to an October 17 changelog, the resulting engine, dubbed DeepSeek Search, already handles hundreds of millions of requests daily and maintains an index containing hundreds of billions of webpages.The New Stack: “the company says its internally developed search engine, dubbed DeepSeek Search, already handles hundreds of millions of requests daily, and maintains an index containing hundreds of billions of webpages. According to its October 17 changelog, the engine supports multilingual searches and uses filtering mechanisms...”
SupportedDeepSeek's $0.75 per 1,000 searches compares with $1 per 1,000 for Parallel's Turbo and Fast services and Perplexity's Fast Search API, $4 for Exa Instant Search, $5 for Brave, and $8 for Tavily. Ceramic.ai charges $0.25 per 1,000 queries on its pay-as-you-go plan.The New Stack: “DeepSeek’s $0.75 per 1,000 successful searches compares favorably with many others, including Parallel’s Turbo and Fast search services and Perplexity’s Fast Search API, both of which charge $1 per 1,000 searches. Meanwhile, Exa charges $4 per 1,000 Instant Search requests, while Brave charges $5 per 1,000 searches. ... Tavily charges $8 per 1,000 basic searches on its pay-as-you-go plan. Ceramic.ai... charges $0.25 per 1,000 queries on its pay-as-you-go plan.”
SupportedMicrosoft retired its Bing Search APIs in August 2025, and Google has closed its Custom Search JSON API to new customers, with existing users required to migrate by January 1, 2027.The New Stack: “Microsoft retired its Bing Search APIs in August 2025, while Google has closed its Custom Search JSON API to new customers, with existing users required to migrate by January 1, 2027.”
SupportedDeepSeek, best known for producing low-cost frontier AI models, announced on Saturday that its Search and Fetch APIs are now available to developers.The New Stack: “DeepSeek made its name with cheap AI models. On Saturday, the company announced it has opened up its internally developed search engine to outside developers through two new APIs, Search and Fetch”
SupportedThe launch was accompanied by DeepSeek-V4.1-Flash-1017, a minor update to the V4.1 Flash model released in September.The New Stack: “The launch also came alongside DeepSeek-V4.1-Flash-1017, a minor update to the V4.1 Flash model the company released in September.”
SupportedLarge language models are trained on data collected up to a fixed point in time, which limits their answers on recent events, updated documentation and other dynamic information. Live web search makes that information available to models, and it is particularly important for AI agents, such as coding agents that need to check the latest software documentation or newly disclosed vulnerabilities.The New Stack: “Large language models (LLMs) are trained on data collected up to a certain point in time, meaning they can’t always answer questions about recent events, updated documentation, or any other kind of dynamic information. By supporting live web search, models can retrieve current information from the internet, which supplements their responses. For developers building AI agents, this is particularly important. A coding agent, for example, might need to consult the latest software documentation, investigate a recently disclosed vulnerability, or otherwise check that a particular library has been u”
SupportedPreviously, developers using DeepSeek's models could connect third-party search services or use DeepSeek's existing search functionality through its model API, which required an additional model request per search and added latency and token costs.The New Stack: “Until now, developers using DeepSeek’s models had several ways to provide this functionality. They could connect their applications to third-party search services, or use DeepSeek’s existing search functionality through its model API. ... accessing DeepSeek’s native search through the Harness required an additional model request, with each search running as a separate call to a DeepSeek model, according to the Harness documentation, introducing extra latency and token costs.”
SupportedDeepSeek has relied on outside providers for much of the past two years.The New Stack: “For much of the past two years, DeepSeek has relied on outside providers to power web search, even as it was building its own.”
SupportedBloomberg reported in January that DeepSeek was recruiting engineers to build a multilingual AI search engine, with plans extending to image and audio search.The New Stack: “In January this year, Bloomberg reported that DeepSeek was recruiting engineers to build a multilingual AI search engine, according to job ads listed by the company. The planned capabilities reportedly extended beyond text searches to include images and audio.”