DeepSeek has quietly released two new APIs — Search and Fetch — that give AI agents and RAG applications direct access to live web content. Search finds web pages for a query and returns structured results including relevance scores, authority tiers, and optional full-page content. Fetch retrieves a single URL and converts it to Markdown. Both use the same bearer-token authentication as DeepSeek's existing model API. The critical detail is that Search is powered by DeepSeek's in-house search engine, not a third-party provider. This is vertical integration: a model company building its own web index rather than depending on Google, Bing, or Brave for grounding data. The search query language is sophisticated — site-specific filtering, date ranges, language restriction, exact-phrase matching, and exclusion operators — suggesting this isn't a thin wrapper over someone else's index. Pricing is straightforward: $0.75 per 1,000 successful requests for both Search and Fetch. Only HTTP 200 responses with actual results are billed; failed requests, empty results, and timeouts cost nothing. Rate limits are 5 QPS per account for each API independently. The count parameter (1-50 results) doesn't affect the per-request price, which means a 50-result search costs the same as a 1-result search. The response schema reveals design priorities. Each search result carries a rerankscore (0-3), an authority tier (0-4), a short snippet for UI display, and a longer summary explicitly described as intended for model consumption. This dual-output pattern — human-readable snippet plus model-readable summary — signals that DeepSeek expects these APIs to be consumed primarily by agents, not by traditional search UIs. There's a notable grace period: tool calls made through DeepSeek Harness, Claude Code, or Codex are currently free, with per-request billing starting November 1, 2026. This is a land-grab strategy — get developers building on the infrastructure before the meter turns on. The mention of Claude Code and Codex as distribution channels is particularly interesting, suggesting DeepSeek is positioning itself as infrastructure for competing model providers' agent ecosystems. The Fetch API is simpler but carries its own implications. It converts any URL to Markdown, subject to the target site's access policies. The 404 error description explicitly acknowledges that sites can block fetching — a nod to the robots.txt and access-control debates that have defined the AI-web relationship. The per-domain QPS limit on Fetch suggests DeepSeek is trying to avoid the aggressive crawling patterns that have poisoned relationships between AI companies and publishers. What we're watching is the infrastructure layer of AI competition solidifying. Search grounding is the oxygen supply for agentic AI. By building its own search engine and offering it as a commodity API at sub-penny pricing, DeepSeek is making a structural bet that whoever controls the retrieval layer controls the agent ecosystem — and that this layer shouldn't belong to Google.
