DeepSeek has added two web tools to its API: one that searches the internet for a query, and one that reads a single page and hands back clean text. Both are described in DeepSeek’s own API documentation, which is a reference page rather than a launch post, so what follows is what the page specifies and nothing more. Both run on the same API key customers already use for DeepSeek’s models.
That detail is the product. Giving a model live web access is one of the first things people want from an agent, and until now it has usually meant a second vendor, a second key and a second bill. A model provider that sells both halves behind one login removes an integration.
On price, the docs are explicit. Search and Fetch each cost $0.75 per 1,000 successful requests, which works out to under a tenth of a cent per call. A request is billed only when it returns a result, so failures, empty searches and timeouts cost nothing. Asking for 50 results costs the same as asking for one. An agent that runs ten searches and reads ten pages to finish a task would spend about 1.5 cents on web access.
There is a free window. Calls made through DeepSeek Harness, Claude Code or Codex already hit these tools, and the docs say they are free until billing starts at midnight Beijing time on 1 November 2026. The page does not say whether direct API use is free in the meantime.
Search takes a query of up to 128 characters and returns between 1 and 50 results, 10 by default. Each result carries a title, a link, a snippet of roughly 100 to 200 characters meant for showing to a person, and a summary of roughly 600 to 800 characters meant for a model to read. Full page text comes back only if the caller asks for it. Results also include a publication date when known, a relevance score from 0 to 3, and a source-authority tier from 0 to 4. The query accepts filters for site, language, exact phrases and date ranges.
The index behind all this is the gap. The docs say Search is powered by DeepSeek’s in-house search engine and stop there. They give no size, no description of how pages are crawled, and no statement about how quickly new pages appear. The relevance score is said to blend relevance, authority and freshness, but how the authority tier is assigned is not explained. An agent’s answers are only as good as what it retrieves, and these scores are DeepSeek’s own judgments.
Fetch takes a web address and returns the page title and its content as Markdown, the plain-text format models read most easily. Returning Markdown instead of raw page code saves the caller a cleaning step. The docs are silent on pages that build their content with JavaScript, on paywalled articles, and on robots.txt. The only guidance is a line saying each fetch depends on what the destination site permits. A 404 error can mean a site blocked the request, the file type is unsupported, or the page held no text to extract. Anyone relying on Fetch for hard sites has to test it.
Rate limits are stated: 100 requests per second for Search and 25 for Fetch, counted per account across all keys, with the two budgets kept separate. Fetch also has a per-domain limit whose size the docs do not give, and no monthly quota appears.
Before 1 November, the free window makes this a cheap moment to run your own 50 hardest URLs through Fetch and count how many come back empty.
Based on DeepSeek’s Search and Fetch API documentation (api-docs.deepseek.com), read on 19 October 2026; the page carries no publication date.
