DeepSeek’s API now does search, link reading and safety checks

DeepSeek says the capabilities behind its consumer app can now be used through its developer API, including web search, file input, link reading, a safety check and voice. In an X post, the company said developers can enable them “with a few parameters”.

The post itself could not be retrieved in full. The details below come from DeepSeek’s API documentation, specifically its chatbot guide and its news page, which carries a banner about the release.

The app’s features become request parameters

According to the chatbot guide, the features work on three endpoints: Chat Completions (/chat/completions), Responses (/responses) and an Anthropic-style Messages endpoint (/anthropic/v1/messages). The only model the guide names for them is deepseek-flash.

Each feature is switched on in the request:

• Web search: a deepseek_web_search tool. With thinking off, the model runs web or image searches and answers with citations. With thinking on, it can search and read links across several reasoning steps.
• URL reading: a deepseek_read_url tool, which fetches pages for links that appear in a user message.
• Safety: a safety_check parameter, for example {"level": "high"}. It reviews input and output and returns a refusal marker (stop_reason: "refusal") when it flags a risk.
• Titles: a title parameter that generates a short conversation title when the answer ends.
• Files and voice: PDF, Word, PowerPoint and image input, plus speech to text and text to speech. These are covered in separate guides that were not detailed on the page we read.

The service runs the searches and page reads itself. The client only displays the returned blocks (thinking, search steps, fetched pages, cited text) and sends them back unchanged in the next round. Usage data includes counters such as web_search_requests and web_fetch_requests.

There are trade-offs. When web tools, high safety or titles are used, the docs say several sampling settings are ignored, among them temperature and top_p. max_tokens still applies, but per inference step rather than to the round’s total output.

DeepSeek also published a browser demo and an open-source repository, deepseek-ai/chat-demo. According to the guide, it covers search, uploads, safety review including images, message editing, read-aloud and voice input. The demo is described as a demonstration only, with the API key and history kept in the browser.

Who can use it, and what is still unclear

Anyone with a DeepSeek API key can try the endpoints, according to the documentation. The pages we read give no pricing for search, URL reading, safety checks, speech or file parsing, and no rate limits or release schedule. The guide points to the Models and Pricing page for costs. It is not clear from the guide whether the new tools are billed per request or per token, so developers should check before building on them.

Nothing here has been independently tested. The behavior described is DeepSeek’s own documentation, not a third-party evaluation.

A bundled stack against pay-per-call search

What changes for developers is how much they no longer have to build. Until now, adding live search, link fetching and moderation to an app on a model API usually meant wiring in separate services. DeepSeek is offering them as server-side tools on the same call, with the citation and refusal handling included.

Rivals sell parts of this separately. Amazon Bedrock AgentCore, for example, added a web search API at $7 per 1,000 queries, a standalone priced product. DeepSeek has not said whether its search is metered the same way. On price, the nearest comparison point is the model side: Baidu priced ERNIE 5.5 Flash at RMB 1 per million input tokens. Chinese labs are competing hard on cheap, fast tiers, and deepseek-flash fits that pattern by name, though its price is not given in these pages.

The guide also carries a compliance warning that Western providers rarely put in a how-to page. Developers who serve the public with the safety check on are told they must handle content review, large model registration and AI labeling themselves. In effect, the safety parameter helps with moderation but does not transfer the legal responsibility to DeepSeek. Operators outside China will need to judge how far that language applies to them.

Signs to watch: published per-call prices, rate limits for the search tool, and whether the same tools reach DeepSeek’s larger models or stay on deepseek-flash. Real-world accuracy of the citations and the safety check will be known only once developers start publishing their own tests.