The reported pricing and end of free server-side calls make retrieval usage part of the operating budget.
PPC Land reports that on 17 October DeepSeek announced Search and Fetch APIs “open to everyone,” priced at $0.75 per 1,000 calls for each endpoint. Search finds web pages for a query and returns structured results; Fetch retrieves one page and returns its content as Markdown. Server-side tool calls from Claude Code, Codex and DeepSeek’s own harness, free today, start costing money at 16:00 UTC on 31 October (midnight 1 November in Beijing). The announcement came as the third post in a thread that began with a model release.
Each search result carries a rerank_score from 0 to 3, an “authority” tier from 0 to 4 (3 or above is “highly authoritative”) and a “summary” of typically 600 to 800 characters that the documentation says is “intended to be read by a model.” PPC Land notes the documentation does not explain how a site earns an authority tier, does not name the fetcher’s user agent or IP ranges, does not say whether robots.txt is consulted, and does not say how the user question or IP address a developer can pass are stored. It reads the summaries as meaning an agent may answer without ever loading the publisher’s page. That is PPC Land’s inference.
In plain English
An “API” lets software use a service directly. A “call” is one request. A “harness” is the program that runs a model with tools. “Markdown” is plain text with simple formatting. A “reranker” orders results by relevance.
Why it matters
Cheap search for agents means more automated visits that never show up as page views, and each site now gets an opaque authority score. Marketers should expect more agent traffic and check what their analytics and bot rules do with it.
Search and retrieval are separate steps in an agent workflow. Search helps identify candidate pages; fetching obtains the content of a selected page. A tool that charges for each step makes the number of calls part of the operating cost, even before the language model processes the results.
For developers, the useful budgeting unit is a completed task. A single answer might involve several searches, multiple page fetches and retries. The headline price per thousand calls is therefore only one ingredient in estimating the cost of a workflow. The end of a free period is a reason to inspect usage, not an estimate of the eventual bill.
For publishers, the reported authority score raises a separate issue. A ranking label supplied to an agent can influence which sources it consults. Without an explanation of how that label is assigned, a site cannot infer a reliable optimization strategy from the number alone.
The reporting also distinguishes documentation gaps from established behavior. An omitted statement about crawling preferences or data handling is not proof that a service ignores those concerns. It means the available documentation does not answer the question. Teams adopting the endpoints should obtain those answers for their use case and check the current pricing and terms before depending on the service.