17 OCT 2026 — DeepSeek is turning web search and page fetching into paid tools for the AI agents built on its models. Two new tools, a Search API backed by its own search engine and a Fetch API that returns a web page as Markdown, cost US$0.75 per 1,000 successful requests each, according to the company's pricing page. Developers using DeepSeek through Claude Code, Codex or DeepSeek's own agent harness have been getting those searches free. That ends on 1 November.
The tools arrived in the same week that DeepSeek's API and chat service went down for about five hours, an outage it compensated by refunding a week of customers' spending.
Disclosure: RECATOOLS is written with the assistance of Claude, made by Anthropic, whose Claude Code is named in this article. Readers should weigh that interest.
What the new tools do
DeepSeek's Search and Fetch guide describes the Search API as "powered by DeepSeek's in-house search engine". It is built for software rather than people: alongside a short snippet for display, each of up to 50 results comes with a longer summary of 600 to 800 characters that the guide says is "intended to be read by a model", and queries of up to 128 characters can use familiar operators such as site: and lang:. Fetch, the companion tool, takes a single URL and hands back the page's content as Markdown.
DeepSeek bills each request rather than each token or result, and only requests that return something count, so empty searches, failures and timeouts cost nothing and asking for 50 results costs the same as asking for one. Each account is limited to five requests per second for each tool, and Fetch has a further limit per website.
Free until November
The pricing page carries the note that matters most for existing users. When the DeepSeek API is used through DeepSeek Harness, Claude Code or Codex, "server-side tool calls already hit the Search and Fetch APIs", the page says. "These calls are currently free of charge; per-request billing starts at 00:00 Beijing time on November 1, 2026."
Anyone running a coding agent on DeepSeek with web search switched on has therefore been using the paid service without being billed. Pandaily, working from DeepSeek's Chinese documentation, reported a two-week free window with the start date to be announced separately. The English pricing page now fixes it at 1 November. An agent that searches dozens of times on every task will start to see those requests on its bill.
The outage
On 15 October, DeepSeek's two API models and its chat service went down in the morning, Beijing time, and were restored at 1.04pm, CGTN reported. In a notice, DeepSeek said it would refund customers' spending from 8 to 14 October as credit, apologised and said it had "reviewed and eliminated single points of dependency". The company did not give a cause. CGTN said DeepSeek's status page listed eight incidents in September.
The uptime record
DeepSeek's status page puts the reliability question in numbers. From July to October it reports 99.74 per cent uptime for the V4 Pro API and 99.47 per cent for the V4.1 Flash API, while the search service shows 99.93 per cent. The Flash figure works out to roughly half a day of downtime over that period.
Where this fits
The new tools add a third pricing dimension to a company already known for unusual billing. DeepSeek charges by the clock, with off-peak rates at half the peak price, and RECATOOLS has tracked its model redirects and billing changes as well as its latest funding round. According to DeepSeek's Chinese documentation, as Pandaily reported, the point is to give agents live web content without a separate third-party search provider.
Check whether web search is switched on in any agent tool pointed at DeepSeek, and watch the 1 November bill. Our LLM cost calculator covers token costs, but search requests are billed on top and need adding separately.
View author profile → · Editorial policy
