DeepSeek hit the spot where rivals park their margin: the per-call meter on tools. No new model, just new math for web-connected agents.
Built-in server-side tools — web search, image search, page fetching — are no longer a separate line on the invoice. Developers pay for model tokens, including the material the model retrieved and processed. TechNode and Pandaily both report the change, effective November 18.
The detail that gets missed: direct Search and Fetch API calls still bill per request. The meter vanished only for search that runs inside the model loop. Call Search yourself and you pay per call, as before.
A separate fee per tool invocation is the standard API pattern, and a comfortable one: every search becomes a billable action. DeepSeek folds it into tokens. For an agent that searches dozens of times per dialogue, what changes is not the model price but the shape of the bill.
There is a second, quieter effect. When in-model search costs nothing extra, a developer is more likely to use the ready-made server tool than to build a retrieval stack and pay for it elsewhere. The message is short: move your tooling inside, stop stitching pipelines.
Retrieved text is now billed as plain tokens. A model that drags three full pages into context costs exactly like a model that drags three full pages into context. Controlling how much context gets swallowed becomes the main way to save money — and the main way to burn it by accident.
Who loses: vendors that sell fresh data as a standalone service. Their pitch was that current information is value in itself. DeepSeek's answer is blunt — current information is just tokens.
Prediction: by the end of the quarter, at least one major API provider will stop billing search as a separate line — or explain publicly why it won't.
Disclaimer: not officially confirmed — no DeepSeek announcement in the cluster, only TechNode and Pandaily.
Source: technode.com · post in Telegram