DeepSeek's Search and Fetch APIs give agents live web access at $0.75 per 1,000 requests

DeepSeek has rolled out two new APIs for connecting AI agents and retrieval pipelines to the live web. The Search API and the Fetch API, described in the company's developer documentation and picked up on Hacker News's front page on October 18, both cost $0.75 per 1,000 successful requests.

How the two endpoints work

Both services are plain JSON APIs served from api.deepseek.com, the same base domain as DeepSeek's model endpoints, and they authenticate with the bearer API key developers already use on the platform. Search posts to /v1/tools/search and Fetch to /v1/tools/fetch.

Search accepts a query of 1 to 128 Unicode characters and returns between 1 and 50 results, defaulting to 10. According to the documentation, it runs on a search engine DeepSeek built in house rather than an outside index — a detail that suggests the company maintains its own crawling and ranking infrastructure alongside its model business.

Results shaped for language models

The response schema leans toward LLM consumers. Every result carries a rerank_score between 0 and 3 that weighs relevance, authority and freshness, with 1.5 and above treated as high quality, plus a separate authority tier from 0 to 4 where a 3 or higher marks a highly authoritative source.

Two text fields serve different audiences: a snippet of roughly 100 to 200 characters meant for on-screen display, and a longer summary of about 600 to 800 characters written for model consumption. Setting the need_content flag to true also pulls full page text into each result.

The query language supports operators familiar from general-purpose search engines: site: and -site: for including or excluding domains, quoted phrases, term exclusions, lang: for language codes, and after: and before: for date bounds. An optional ori_question field passes along the user's original question to help the engine interpret a rewritten query, and a localization parameter takes an IP address to tailor results by region.

Fetch converts pages to Markdown

Fetch takes a single URL — at most 8,192 characters, HTTP or HTTPS, with no embedded username or password — and returns the page title alongside its content converted to Markdown. Retrieval depends on the destination site being reachable and willing to serve the request; when no usable Markdown can be extracted, the call fails with a 404. Errors across both APIs follow the OpenAI-compatible error format.

Charging only for successes

Billing is per successful request rather than per token or per result. A call counts as billable only when it returns HTTP 200 and actually produces something: a search must return a non-empty results array, and a fetch must return page content. Empty result sets, failures and timeouts cost nothing, and the count parameter has no effect on the price.

The documentation also notes that server-side tool calls made through DeepSeek Harness, Claude Code and Codex already route through these two APIs and remain free of charge for now.

Why it matters

Live web access has become a baseline requirement for agent frameworks and RAG systems, and DeepSeek is now selling that plumbing directly alongside its models, putting it in competition with established search API vendors on price. The success-only billing model lowers the barrier to adoption, since developers pay nothing for empty or failed lookups. Just as telling is the shape of the API itself: fields like long-form summaries, rerank scores and authority tiers suggest a product designed from the start for machine readers rather than retrofitted from a consumer search box. For teams building agents on DeepSeek models, the company now supplies the retrieval layer too.