DeepSeek has opened its own search engine to outside developers. Two endpoints arrived: Search finds web pages and returns structured results, Fetch pulls a single page and hands it back as Markdown. Both shipped over the weekend alongside V4.1-Flash-1017, which I wrote about yesterday. By DeepSeek’s own announcement, the engine behind it serves hundreds of millions of requests a day.
Two endpoints, one key
Both sit under https://api.deepseek.com and take the same bearer token as the models: POST /v1/tools/search and POST /v1/tools/fetch.
Search takes a query of 1 to 128 characters and up to 50 results (count, default 10). Three optional fields go with it: ori_question for the user’s original question, to help the engine understand the query, localization for an IP address to localize results, and need_content, which returns page content instead of metadata alone. The API also understands operators: site: and -site:, exclusions with -term, exact phrases in quotes, lang:zh,en, and date bounds with after: and before:.
Fetch takes a single url of at most 8,192 characters and returns a title plus the content, the latter as Markdown.
What one hit carries
The per-result payload is the interesting part. Next to title, URL and a short snippet of 100 to 200 characters meant for display sits a summary of 600 to 800 characters, explicitly meant for models. Then two scores: rerank_score from 0 to 3, where DeepSeek calls 1.5 and up high quality, and authority from 0 to 4, where 3 and up counts as highly authoritative. A missing value shows as -1.
Anyone filtering sources programmatically gets two numbers there that usually have to be built in-house.
5 requests per second, per account
The limits are tight: 5 QPS for Search, 5 QPS for Fetch, and for Fetch an extra per-domain cap on top. Counting happens per account no matter how many API keys are involved, separately from the model concurrency limits. Past that you get HTTP 429.
Pricing only counts success. The 75 cents per 1,000 requests applies to Search and Fetch separately, and a call is billable only when it returns HTTP 200 with a result — a non-empty results array for Search, actual page content for Fetch. Failures, empty result lists and timeouts are free, and asking for ten hits versus fifty makes no difference to the price.
Until November 1, nobody bills you inside the Claude Code harness
The line that matters for Claude users sits in the billing section: server-side tool calls made through DeepSeek Harness, Claude Code or Codex are free for now. Per-request billing for those starts on November 1, 2026, at midnight Beijing time.
That’s two weeks to find out how someone else’s web search feels inside a Claude Code workflow, at no cost. The direction is worth noting: a Chinese lab is marketing its infrastructure with a competitor’s harness by name. The same pattern showed up in the V4.1-Flash-1017 model card, where DeepSeek measured its agent scores in the Claude Code harness. Whoever sells tools goes where the work happens.