Summary: DeepSeek has opened the search engine behind its own products as two raw JSON APIs — Search and Fetch — under the same bearer auth as its model API. On October 17, 2026, the company announced the public availability of its in-house engine, which it says indexes hundreds of billions of pages and filters SEO spam and low-quality AI-generated content. For builders wiring agents to the live web, the interesting parts are the pricing model and the response shape, not just the brand.
What you actually get
The Search & Fetch guide describes two endpoints, both POST requests against api.deepseek.com:
POST /v1/tools/search— takes a query, returns aresultsarray with title, URL, snippet, summary, and optionally full page content.POST /v1/tools/fetch— takes a URL, returns the page as Markdown under acontentfield.
The response design is clearly built for LLM consumption rather than UI display. Each result carries both a snippet (100–200 characters, meant for humans) and a summary (600–800 characters, meant for a model), plus two quality signals: rerank_score from 0 to 3, where 1.5 and above counts as high quality, and an authority tier from 0 to 4. That lets you cheaply filter before paying attention to content — or skip need_content entirely and work from summaries.
The query syntax is passed through intact, so you can compose operators directly in the query string:
{
"query": "DeepSeek site:github.com OR site:huggingface.co lang:en after:2025",
"count": 10
}Supported operators include site:, -site:, phrase exclusion, lang:, and after:/before: date filters. If you’ve built agent search workflows before, this is the difference between hoping the engine understood your intent and programming the retrieval precisely.
Billing that rewards good queries
Both APIs cost $0.75 per 1,000 successful requests. A request is only billable when it returns HTTP 200 with an actual result: empty search result sets, failed fetches, and timeouts are free. count doesn’t affect price, so asking for 50 results costs the same as asking for 1.
This is a different contract from token-metered search APIs, and it changes incentives: a badly worded query that returns nothing costs you nothing but latency, while a query that returns junk costs full price. The per-request model also pairs naturally with the earlier point about filtering on rerank_score — you’re paying per retrieval decision, so the quality signals are doing real budget work.
Rate limits and failure modes
Each API is capped at 5 QPS per account, calculated across all API keys and independent of model concurrency. Fetch has an additional per-domain limit on top of that. Errors follow the OpenAI-compatible format; the docs advise retrying only on 408, 429, and 5xx, and topping up on 402 before retrying.
Fetch is also honest about its limits: it’s subject to the target site’s availability and access policies, and a 404 can mean anything from a blocked URL to a page with no extractable Markdown. If your agent pipeline treats fetch as guaranteed, it shouldn’t.
One timing note: if you use DeepSeek through Harness, Claude Code, or Codex, server-side tool calls already hit these APIs and are currently free — per-request billing for those starts at 00:00 Beijing time on November 1, 2026. Budget accordingly.
Where this fits in the stack
The change log positions DeepSeek Search as an AI-native engine that filters GEO manipulation and low-quality AIGC, which is a real concern for agents that ground answers in web content. This is the same consolidation trend we’ve covered before — OpenRouter shipping consistent search and fetch across every model rather than leaving each provider to improvise. DeepSeek doing it in-house means one vendor now owns the model, the search engine, and the fetch layer, with a single API key.
The pragmatic next step if you’re evaluating: run the same query set against this and your current search provider, and compare the rerank_score distribution and empty-result rate. Those two numbers will tell you more than any benchmark about whether $0.75 per 1,000 is cheap for you.
Sources
View the original post on X