DeepSeek opens search and page fetching in its API: here is what it costs

DeepSeek now offers two tools that developers can use to give an AI model access to the web. An API is a digital gateway through which a program can ask a service to perform a task. In this case, the tasks are searching and fetching web pages.

The two endpoints are hosted at api.deepseek.com. The first, /v1/tools/search, finds web pages through DeepSeek's own search engine. The second, /v1/tools/fetch, retrieves a single web page and converts it to Markdown, a simple text format that is easy for an AI model to read. Both require a bearer token, in other words an access key.

Pricing and free period

Both APIs cost USD 0.75 per 1,000 successful requests. You only pay for responses with content (HTTP 200). Errors and empty search results are free.

Per-request billing for server-side tool calls via DeepSeek Harness, Claude Code or Codex begins at 00:00 China time on November 1, 2026. Until then, they are free. The documentation does not state whether direct calls to the API are also free during the period.

Limits and capabilities

According to DeepSeek's documentation, the account-level limit is 100 requests per second for search and 25 per second for page fetching. Fetching also has a per-domain limit, but the value has not been published. If you exceed the limits, you get an HTTP 429 error.

Search supports operators such as site:, lang:, after: and before:, as well as exact phrase search. According to the independent review on mer.vin, search queries are limited to 128 characters. The results include a relevance score from 0 to 3, an authority level from 0 to 4, an excerpt and a summary of typically 600-800 characters.

Thinking mode requires care

DeepSeek's thinking mode supports tool calls. The model can go through several rounds of reasoning and tool calls before it gives its final answer. If a request includes the tools parameter, reasoning_content from earlier rounds must be sent back in subsequent requests. Otherwise, the API responds with a 400 error.

Caveats

Mer.vin recommends running an evaluation against your current provider before switching. The reason is that DeepSeek has not published benchmarks, an SLA (agreed uptime guarantee) or a data retention policy.