DeepSeek Starts Charging for Web Search in Its API
DeepSeek, the Hangzhou-based AI lab whose open models made it one of China’s best-known AI names, has quietly put a price on something its users had come to take for granted: looking things up on the web. In an announcement on October 18, the company said its API now supports web search and page fetching, charged by the request rather than by tokens. The rate is 5 yuan (roughly 70 US cents) per 1,000 successful requests.
The distinction that matters for developers is what counts as successful. DeepSeek’s notice frames the fee around completed calls to its Search and Fetch interfaces, the two server-side tools that let a model pull in fresh information from outside its training data. In other words, customers pay when the tool actually returns something, not merely when it is asked.
The company also explained where these charges may already be showing up. When users run DeepSeek’s models inside agent-style tools such as DeepSeek Harness, Anthropic’s Claude Code or OpenAI’s Codex, the software can reach these Search and Fetch endpoints through server-side tool calls. Until now, that plumbing was largely invisible to end users; the new pricing makes it a line item.
Web access has become a standard feature of the AI agent era. A model that can only answer from its training data goes stale, while one that can search, open pages and cite them can handle research, shopping comparisons and coding questions that depend on current documentation. Running that infrastructure is not free: every query fans out to crawlers, page parsers and content filters, and the costs scale with usage rather than with the size of the model.
For DeepSeek, metering search is also a business signal. The company built its reputation on releasing strong models openly and charging unusually little for API access, a strategy that won it global attention and a large developer following but left open the question of how it would earn money. Paid tool calls offer a revenue stream that grows with how heavily customers lean on agents, without raising the headline price of the models themselves.
For developers, the calculus is straightforward but worth doing. A chat assistant that occasionally searches the web will barely notice the change. An agent that fires dozens of lookups per task, however, could see search fees rival its token bill. Teams building on DeepSeek will need to decide when a model should search and when it should simply answer from what it already knows.
The pricing also lands in a crowded field. Search-augmented answers are now offered by chatbot apps, browser makers and cloud platforms worldwide, and several competitors bundle retrieval into their API rates rather than billing it separately. DeepSeek’s per-request model is transparent, but it asks customers to think about a cost that most AI services still hide.
What the announcement does not change is the underlying pitch: capable models at low prices, accessible to developers far beyond Silicon Valley. Charging for web search is less a reversal of that philosophy than an attempt to make it sustainable, one successful request at a time.