An agent that reads the web used to carry two meters: one for the words the model produced and one for every page it opened. DeepSeek has removed the second meter from model-driven calls. The standalone search API still has one.
The change took effect at noon Beijing time on November 18, 2026. When a developer calls a DeepSeek model and the model decides to search the web or fetch a page, that action is no longer billed per request. Only the model's token consumption appears on the invoice.
What changed, precisely
DeepSeek's own site, as reported by several technology outlets, states that server-side Search and Fetch triggered during model calls are no longer charged separately. Three triggers are named: the built-in web tool (联网工具), the web_search tool in the Anthropic API, and the web_search tool in the Responses API.
This is a narrow and specific change. It applies to tool use that happens inside a model call, initiated by the model. It does not make the internet free, and it does not abolish all search charges.
The standalone interfaces are a different product. A developer who calls the Search API or Fetch API directly still pays per successful request, at five yuan per thousand (≈ US$0.69 / HK$5.4, approximate). That price was confirmed on the official pricing page in reports published after the change.
So the new rule is best read as a distinction between two routes to the same capability. Let the model decide to search during a conversation, and the search is bundled. Call search as a service in your own code, and you still pay for it.
The fee that lasted 18 days
The per-request mechanism being retired was barely a month old. On October 17, DeepSeek released the Search and Fetch interfaces alongside its V4.1-Flash-1017 model, using a self-developed search engine and pricing both at five yuan per thousand successful requests. A two-week grace period meant tool calls through DeepSeek Harness, Claude Code and Codex were initially free. Formal per-call billing began on November 1.
It ended on November 18, an eighteen-day run. That is a short life for a pricing scheme, and it says something about how quickly developer feedback travels when a tool is embedded in agent workflows.
For an agent, search is not an occasional add-on. A research task may trigger dozens of queries and page fetches before producing a paragraph. A per-call fee turns every extra step into a small financial decision, which is exactly the kind of friction that makes an agent feel unreliable. Moving the cost into tokens does not make it disappear, but it makes the bill predictable.
The toolchain adjusted too
DeepSeek Harness, the company's own agent harness, published a pre-release update numbered v0.2.2-rc.5 to reflect the new rules. According to its GitHub changelog, when the harness runs on a DeepSeek model it uses DeepSeek's server-side search, which carries no per-call charge. When it runs on another model, it defaults to the standalone Search API, which still does.
That split is important for anyone building a multi-model application. The economics of browsing now depend on which model is in the loop, not just on how many pages are read.
Cloud providers moved as well. Tencent Cloud issued a notice on November 20 saying that for DeepSeek models marked as supplied directly by the original manufacturer, server-side Search and Fetch would no longer be billed from November 18. It also said charges incurred after that time would be refunded, with refunds completed by November 30.
Why now
The timing is hard to read as a simple price cut. Two days before the billing change, on November 16, some DeepSeek users reported that long conversations had stopped, input limits had shrunk and daily quotas had tightened. Some messages attributed the interruption to insufficient computing power. Users asked publicly for a paid tier that would preserve their previous experience.
Against that backdrop, removing a charge looks less like generosity and more like a decision about where to compete. DeepSeek appears to be making browsing cheaper for developers while managing consumer demand separately. The two moves are not contradictory, but they describe a company under real capacity pressure.
Honest limitations
I could not open DeepSeek's official pricing page or API documentation directly. The effective date, trigger list, retained standalone price and refund schedule are taken from reports quoting those pages, including Tencent News, Phoenix Technology, The Paper and Xinhua Finance. Treat them as well-corroborated secondary reporting rather than a first-hand reading of the terms.
The eighteen-day figure counts from November 1, when formal per-call billing began, to November 18. The interfaces themselves launched on October 17. Different outlets describe the earlier grace period slightly differently, so the exact duration depends on which starting point is used.
No detail is given here about how search costs are reflected in token usage, because the company did not publish a clear explanation in the material reviewed. Developers should inspect their own invoices before assuming the total cost has fallen.
What readers can do now
- If you build agents, check which search route your framework uses. The same browsing task can be bundled or billed per call depending on the model and the tool configuration.
- If you were charged for server-side search after November 18, look for a refund. Cloud providers announced refund windows, and the money will not return automatically to every account.
- If you use the standalone Search or Fetch API directly, do not assume it is free. That price was still listed in the official pricing material after the change.
