DeepSeek has documented Search and Fetch APIs that give agents and retrieval-augmented generation applications access to current web material. Their defining commercial mechanism is not token-based pricing: a call becomes billable only when it succeeds and returns a result or page content, according to the DeepSeek Search and Fetch guide.
For commercial operations teams, this creates a specific option for workflows that need recent public information. It also creates a new unit to monitor. The cost driver is the number of productive tool calls, not the length of the retrieved page or the number of search results requested.
DeepSeek separates discovery from page retrieval
Search and Fetch perform different stages of web research. Search discovers relevant pages and returns a structured results array. Fetch retrieves content from a page. Both expose raw JSON interfaces and use the same bearer authentication as the wider DeepSeek API, the documentation states.
This separation matters when designing an agent. The agent does not need to retrieve every page discovered by a query. It can inspect search results, select relevant sources and invoke Fetch only for pages needed to complete the task. That sequence gives the application control over how many retrieval calls it makes.
The tools are read-only. They supply web material to an agent or RAG pipeline but do not submit forms, change records or perform actions on the retrieved site. Their role is therefore evidence collection rather than browser execution.
Conceptual illustration of Search discovering pages and Fetch retrieving selected content; not a photograph of the service.
Billing follows usable responses, not token volume
DeepSeek bills Search and Fetch per successful request rather than by tokens or by the volume of results returned. A Search call is billable when it receives a successful response with a non-empty results array. A Fetch call is billable when it receives a successful response containing page content. Failed calls, timeouts, searches without matches and fetches without content are excluded, according to the billing rules in the guide.
The result-count setting does not change the per-request price. Requesting more returned matches therefore does not create a different charge for that individual successful search, although an agent that subsequently fetches more selected pages will generate more successful retrieval calls.
This differs from budgeting a model interaction through input and output tokens. A commercial team needs to count successful searches and successful page retrievals at the workflow level. Repeated searches, broad query expansion and fetching every candidate source become the relevant design choices.
DeepSeek also documents distinct error handling. Calls receiving HTTP 408, 429 or a 5xx response should be retried after waiting, while HTTP 402 requires an account top-up. Other 4xx responses should not be repeated without changing the request, according to the same technical guide. This prevents an unchanged invalid request from entering a pointless retry loop.
Conceptual illustration of result-based request billing; not a photograph of DeepSeek infrastructure.
Search operators pass through unchanged
The Search API forwards a query while preserving its search syntax. DeepSeek’s examples combine a search term with operators, and the guide says multiple operators can be placed in the same query. That lets an application narrow discovery before requesting full page content from Fetch, as described in the API documentation.
For a commercial research task, this makes query construction part of the workflow rather than an incidental prompt detail. Domain restrictions or other supported search operators can focus discovery on the intended source set. The agent still needs rules for selecting results and handling an empty response; preserved syntax does not establish that the returned pages are accurate, authoritative or relevant.
The published guide explains request behavior and billing conditions, but it does not provide comparative evidence about ranking quality, web coverage or retrieval latency. Those characteristics remain task-specific evaluation points rather than documented advantages.
A bounded account-research workflow fits the mechanism
Consider a hypothetical sales-research agent preparing a briefing about a prospective customer. Search would discover recent public pages matching a constrained query. The application would then select sources based on predefined relevance criteria and use Fetch to retrieve only the chosen pages. The language model could summarise that retrieved material into the briefing while retaining source references.
The mechanism suits this bounded task because discovery and retrieval remain separately observable. The team can record which query produced the candidates, which pages were fetched and how many content-bearing calls completed. A broad instruction to “research the company” offers less control because the agent can expand searches and retrievals without a clear stopping condition.
DeepSeek says server-side search and fetch calls made through DeepSeek Harness, Claude Code or Codex are currently free, with per-request billing beginning at 00:00 Beijing time on November 1, 2026, according to the announcement in its guide. Teams already using those integrations therefore need to identify which workflows invoke the tools before that billing boundary.
The immediate decision is whether live-web access solves a defined research task whose successful calls can be measured. Search and Fetch provide the retrieval mechanism and a result-based billing rule; they do not establish source quality or the business value of the resulting brief. Those outcomes still require testing against the exact commercial workflow.
