DeepSeek has opened its in-house search engine to external developers through a new Search API, according to the company's official changelog entry dated October 17. The move exposes infrastructure that previously powered only DeepSeek's own chat and API search features, and positions the Chinese AI company in direct competition with the web-search layers offered by OpenAI, Anthropic and Perplexity.
The company describes DeepSeek Search as an AI-native general-purpose search engine that has indexed hundreds of billions of web pages, supports multilingual queries, and returns results in real time. It says the engine applies detection mechanisms to filter SEO spam, GEO manipulation and low-quality AI-generated content, an increasingly common concern for model builders whose training and retrieval pipelines are flooded with machine-written pages. The company also says its search backend already handles hundreds of millions of search requests per day, a figure that is company-reported and not independently verified. A companion Fetch API is documented in the same developer guide.
The same changelog shipped a model refresh. The deepseek-flash model name now routes to DeepSeek-V4.1-Flash-1017, with no change to API pricing or calling conventions, and the older deepseek-v4-flash and deepseek-v4-flash-vision-exp names remain accepted even though those models have been retired, with their requests served by the new version. DeepSeek recommends version 0.2.1-rc.1 or later of its own DeepSeek Harness agent framework for the best experience.
Benchmark gains are vendor-reported. Against the prior V4.1-Flash release, the company lists scores including 42.6 on HLE without tools (44.0 on its text subset), 42.3 on Terminal-Bench 4.0, 77.0 on DeepSWE v1.1 and 68.2 on NL2Repo-Bench, alongside results on a dozen other agentic and multimodal benchmarks. The company says the update reduces task runtime and output length to reach what it calls a performance-cost Pareto optimum, and that the model was specifically tuned for agent frameworks including DeepSeek Harness, Tencent's CodeBuddy and WorkBuddy.
The strategic logic of opening the search API is straightforward. Retrieval quality has become a differentiator for agentic systems, where an agent's answers are only as good as the pages it can find and parse. By selling the same index that serves its own assistant, DeepSeek gives agent developers a reason to build on its stack beyond model price, and monetizes an asset most frontier labs have kept in-house.
The timing is notable too. DeepSeek's API business has been under strain from repeated outages this year, and the company's earlier moves, from open-sourcing its training stack for domestic chips to putting its agent harness on the desktop, have all pushed in the same direction: own the full path from silicon to search to agent. The Search API extends that ownership one layer outward, to the open web itself.
For developers, the practical questions are the ones the changelog did not answer: standalone pricing for search calls, rate limits, and how the filtering layer treats borderline content. Those details live in the Search & Fetch guide rather than the announcement, and will shape whether the API becomes a default retrieval layer for the agent ecosystem or just another option in a crowded field.
Comments (0)
Log in to join the discussion
Log InNo comments yet