DeepSeek sells its search engine to agents: $0.75 per thousand calls, results come with a summary written for a model, not a person
DeepSeek opened its in-house search engine as two raw JSON endpoints on October 17, and HN found it on Sunday morning and pushed it to 507 points and 249 comments. The endpoints are POST /v1/tools/search and POST /v1/tools/fetch on api.deepseek.com, same bearer key as the chat API. Search costs $0.75 per 1,000 successful requests, fetch costs the same, and nothing is charged for failed calls. The change log describes the engine behind it as serving hundreds of millions of queries a day over hundreds of billions of indexed pages, with filters against SEO spam, GEO manipulation and low-quality AI-generated content. That is the search engine DeepSeek built for its own models, now sold by the call.
The schema is what makes this an agent product rather than another Bing wrapper. Each result carries a snippet of 100 to 200 characters for display in a UI, and separately a summary of 600 to 800 characters that the docs say is intended to be read by a model. There is a rerank_score from 0 to 3 with 1.5 as the stated bar for high quality, an authority tier from 0 to 4, a date field in RFC 3339 or null, and an optional full content field if you set need_content. Query operators are the usual site:, -site:, exact phrase, minus phrase, lang:zh,en and after:/before:. Count goes up to 50 and there is an ori_question field so you can pass the user's actual question alongside the rewritten query. Fetch returns a title and the page as Markdown. Errors follow the OpenAI-compatible format, which tells you who the docs are written for.
The pricing page has the line that matters for anyone already using DeepSeek inside a coding agent. When you call the DeepSeek API through DeepSeek Harness, Claude Code or Codex, the server-side tool calls already go through these two endpoints, and those calls are free until per-request billing starts at midnight Beijing time on November 1. So the web search that V4.1-Flash has been doing inside your terminal for the past two weeks was this engine the whole time, and the price of it was zero until now. The quick-start page also says DeepSeek Harness is in developer preview for harness builders worldwide, though the guide it links to was not reachable from here.
Why a lab would do this is the interesting part. The same week The Verge wrote that websites are closing their doors to agents and Schwab and JPMorgan said they block them, DeepSeek put a price on the layer that every agent depends on and nobody owns. Exa, Tavily, Brave and Perplexity have been selling agent search for a year, but they are search companies. DeepSeek is a model lab selling its retrieval index as a standalone utility at a price that undercuts most of them, with a summary field tuned to its own models. The strategy reads like the Flash pricing applied to search: make the per-call cost small enough that nobody bothers to build their own crawler, and let the index become the default input for the open-weight agent stack. The HN thread was mostly people pasting their first results and arguing about whether the authority tiers are any good, which is the right question and one nobody has answered yet; there is no public evaluation of the index against Exa or Brave as of Sunday night.
Links: https://api-docs.deepseek.com/guides/search/ , https://api-docs.deepseek.com/updates , https://api-docs.deepseek.com/quick_start/pricing/ , https://news.ycombinator.com/item?id=50138531
← Back to all articles
The schema is what makes this an agent product rather than another Bing wrapper. Each result carries a snippet of 100 to 200 characters for display in a UI, and separately a summary of 600 to 800 characters that the docs say is intended to be read by a model. There is a rerank_score from 0 to 3 with 1.5 as the stated bar for high quality, an authority tier from 0 to 4, a date field in RFC 3339 or null, and an optional full content field if you set need_content. Query operators are the usual site:, -site:, exact phrase, minus phrase, lang:zh,en and after:/before:. Count goes up to 50 and there is an ori_question field so you can pass the user's actual question alongside the rewritten query. Fetch returns a title and the page as Markdown. Errors follow the OpenAI-compatible format, which tells you who the docs are written for.
The pricing page has the line that matters for anyone already using DeepSeek inside a coding agent. When you call the DeepSeek API through DeepSeek Harness, Claude Code or Codex, the server-side tool calls already go through these two endpoints, and those calls are free until per-request billing starts at midnight Beijing time on November 1. So the web search that V4.1-Flash has been doing inside your terminal for the past two weeks was this engine the whole time, and the price of it was zero until now. The quick-start page also says DeepSeek Harness is in developer preview for harness builders worldwide, though the guide it links to was not reachable from here.
Why a lab would do this is the interesting part. The same week The Verge wrote that websites are closing their doors to agents and Schwab and JPMorgan said they block them, DeepSeek put a price on the layer that every agent depends on and nobody owns. Exa, Tavily, Brave and Perplexity have been selling agent search for a year, but they are search companies. DeepSeek is a model lab selling its retrieval index as a standalone utility at a price that undercuts most of them, with a summary field tuned to its own models. The strategy reads like the Flash pricing applied to search: make the per-call cost small enough that nobody bothers to build their own crawler, and let the index become the default input for the open-weight agent stack. The HN thread was mostly people pasting their first results and arguing about whether the authority tiers are any good, which is the right question and one nobody has answered yet; there is no public evaluation of the index against Exa or Brave as of Sunday night.
Links: https://api-docs.deepseek.com/guides/search/ , https://api-docs.deepseek.com/updates , https://api-docs.deepseek.com/quick_start/pricing/ , https://news.ycombinator.com/item?id=50138531
Comments