DeepSeek Flash 1017 Ships Search and Fetch APIs, Dropping the Barrier to Agent Web Search
Agent developers who want models to search and read the live web get a hosted Search and Fetch API at 5 yuan per 1000 successful requests, replacing hand-built scrapers and result processors. At that price, checking several sources per answer stops being a cost concern.
DeepSeek Flash 1017 opens two APIs that give agent developers live web grounding. Search finds relevant material and Fetch pulls the page body text, so an agent can locate content, read the full context, then compose an answer. A developer who previously ran their own search tooling reports the switch removed a chunk of tedious work: search result processing and page extraction no longer need maintenance.
Pricing sits at 5 yuan per 1000 successful requests per API. One search plus three page reads runs about two cents, before model invocation fees. At that rate, repeated search rounds and cross-checking sources become affordable inside agent loops.
The same developer flags a rumor making the rounds: the search layer is not Baidu-backed. DeepSeek's official announcement says the search engine is self-developed and has been used long-term, so the quality gains should not be credited to a Baidu partnership.
Search quality is the hard part of web grounding, not just the plumbing: the developer's own implementation worked but underperformed, while the official APIs produced noticeably better results without extra tuning.
The pricing math means agent loops can afford many searches and fetches per task, which changes how aggressively a developer can verify sources before answering.
The need to publicly correct the Baidu rumor shows how quickly assumptions about search backends spread, and official statements are the only reliable source.