DeepSeek has quietly made its built-in web search free at the tool level. Its Models & Pricing page now says that from 12:00 Beijing time on 18 November 2026 (04:00 GMT), “the server-side search and web fetch tool calls triggered when calling the model are no longer billed”. The results those tools return are appended to the model context and charged as ordinary token usage. The Chinese version of the page carries the same notice.
That reverses what we reported only three days ago. When we covered DeepSeek building search, deep thinking search and URL reading into its chat API, those built-in tools were priced the same as the standalone APIs: $0.75 per 1,000 searches or fetches, plus tokens. For any request where the model decides to search, that per-call fee is now zero.
What is free and what is not
DeepSeek’s Search & Fetch guide sets out the split clearly:
- Free (tokens only): text search, image search and page fetch run by the model through server-side tools. That covers the web tools (
deepseek_web_search,deepseek_read_url), theweb_searchtool in the Responses API and the Anthropic-compatible API, Web Search in Claude Code, and search in DeepSeek Harness when the “Web search provider” is set to DeepSeek server-side search or Auto, which is the default. - Still billed: direct client calls to the standalone Search and Fetch APIs, at $0.75 per 1,000 successful requests each (¥5 on the Chinese page). DeepSeek gives two examples: Harness set to use the DeepSeek Search API as its provider, and Codex Alpha Search.
The standalone rules are unchanged: only HTTP 200 calls that return results are billed, and failed calls, empty searches and timeouts are free.
“Free” still costs tokens
The change does not make grounded answers free. Whatever the tools fetch is added to the model’s context, and DeepSeek charges for every token of it. On deepseek-flash that is $0.30 per million cache-miss input tokens at peak and $0.15 off-peak, according to the same pricing page. In deep thinking search the model can make “up to about 20 tool calls before it answers”. Under the old pricing that alone could add 20 × $0.00075 = $0.015 to one request. Now the bill depends entirely on how much page text the model pulls in.
The incentive has therefore moved. DeepSeek no longer earns per lookup on model-driven search, so it gains nothing from extra searches. It earns on tokens, so it benefits when the model reads more. For developers the sensible default is now to let the model search and keep an eye on context size, rather than rationing searches.
Why it matters for search visibility
This is the third move in a month that makes DeepSeek’s in-house index more widely used. In October it opened the index as an API (its change log claims “hundreds of billions of web pages” and filtering of “SEO spam, GEO manipulation, and low-quality AIGC”). On 13 November it built that index into the chat API. Now model-triggered use costs nothing at the tool level. Each step lowers the cost for a Chinese app, chatbot or agent of grounding its answers in DeepSeek’s results rather than Baidu’s or Tencent’s.
The Harness detail matters most. When we covered DeepSeek Harness passing 1.54 million daily users, agent search calls were being billed. With the default provider now “Auto” and server-side search free, the cheapest option for most Harness users is DeepSeek’s own index, and few will change the setting.
All this comes while DeepSeek is rationing its free consumer app. Caixin reported on 18 November that since 15 November long conversations inactive for 72 hours can no longer be continued, and that long documents are capped at the first 20 pages or 10,000 tokens, both because of a shortage of computing power. We have since updated our 17 November story with those details. Caixin quotes Haitong International’s Yao Shuqiao as saying DeepSeek has to give paying customers priority. That is clearly what is happening here: free users get less, and paying API users get search at no extra charge.
My take
This is a land grab. Baidu sells its search API by the call, and we wrote about Alibaba cutting its web search API price alongside Qwen4-Max. DeepSeek has now taken the per-call fee off model-driven search altogether. If you are building a Chinese-language agent, that makes the decision easy.
For SEOs targeting China, the “DeepSeek index” is becoming a retrieval layer under a growing number of products that never show the DeepSeek name. What you can control hasn’t changed. Make sure your pages can be fetched (DeepSeek still hasn’t published a crawler user agent, so be careful with blanket bot blocking). Put clear titles and machine-readable dates on every page. Put the key facts high on the page, because a model paying per token has every reason to read as little as it needs.





Leave a Reply