DeepSeek’s search API caps every account at 5 requests a second

Five a second. That's the whole budget. DeepSeek's new Search API will happily sell you a ranked list of up to 50 web pages for three quarters of a cent per thousand calls, but its own guide says the account gets 5 queries per second on Search and 5 on Fetch, and it doesn't matter how many keys you mint. We read the Search & Fetch guide line by line, because the press coverage has mostly stopped at the price.

What one call buys

Two endpoints, both POST, both behind the same bearer key as the chat API: https://api.deepseek.com/v1/tools/search and https://api.deepseek.com/v1/tools/fetch. Search takes a query of 1 to 128 Unicode characters and a count between 1 and 50. Set need_content to true and each result carries the page text in a content field, next to a 100 to 200 character snippet meant for a UI and a 600 to 800 character summary meant for a model. Each result also gets a rerank_score from 0 to 3 (DeepSeek calls 1.5 and up high quality) and an authority tier from 0 to 4, with -1 for unknown.

The query operators are the familiar ones: site:, -site:, exact phrases, lang:zh,en, plus after: and before: for dates. Fetch is simpler. You send a URL of up to 8,192 characters with no username or password in it, and you get back a title and the page as Markdown.

Here's the bit we'd build around. Billing is per successful request, and the guide says plainly that count "does not affect the per-request price". So a single search with count: 50 and need_content: true costs the same $0.00075 as a search returning one link. For a RAG pipeline that today does one search and then fetches the top ten pages, that's eleven billable calls collapsing into one. The catch: the guide doesn't say how long content can get or whether it's truncated, and we haven't tested it. If it's clipped, you'll still need Fetch for the pages that matter.

Failed calls are free. A request bills only on HTTP 200 with a non-empty result array (or, for Fetch, actual page content), so an empty search costs nothing, and neither does the 404 Fetch returns for a blocked or unextractable page. Timeouts aren't billed either.

The 5 per second ceiling, worked out

The limit is set "at the account level, regardless of which API key is used". Search and Fetch are counted separately, and neither touches your model concurrency. Run the arithmetic and one account tops out at 432,000 searches a day, which at list price is $324. That's not a lot of headroom for a team. If ten developers each run an agent that fans out three searches at once, you're over the line on the first burst and eating HTTP 429s.

Fetch is tighter than it looks. On top of the account's 5 per second there's a per-domain limit, and DeepSeek doesn't publish the number. An agent that tries to pull twenty pages from one documentation site in a loop will hit it, and the docs only tell you to wait and retry on 429. We'd put a small per-host queue in front of Fetch from day one rather than discover the limit in production.

For scale, Anthropic's own web search tool lists $10 per 1,000 searches, so DeepSeek is about 13 times cheaper per call. They aren't the same product, though. Anthropic's runs inside the model turn and returns encrypted results with citations, while DeepSeek hands you raw JSON to feed to whatever model you like. Cheap per call, narrow per second: that's the trade.

One field gave us pause. localization takes "an optional IP address" to localize results. If you pass your end user's IP straight through, you're sending personal data to DeepSeek, and the guide says nothing about how long it's kept. We'd send a coarse, server-side value or leave it empty until there's a written policy. I might be overcautious here, but it's an easy field to wire up without thinking.

The free part ends on 31 October

There's a note at the bottom of the billing section that's easy to skip. When you use the DeepSeek API through DeepSeek Harness, Claude Code or Codex, "server-side tool calls already hit the Search and Fetch APIs". Those calls are free for now. Billing starts at 00:00 Beijing time on 1 November 2026, which is 16:00 UTC on 31 October, or noon in New York.

So if your team points a coding agent at DeepSeek's Anthropic-compatible endpoint, its lookups already run on this search engine, without a line item. From that afternoon on, every lookup the agent decides to make lands on your bill and inside the same 5 per second budget as your own scripts. The guide doesn't describe a switch to turn those server-side calls off, and we didn't find one. Our read is that direct calls to /v1/tools/search bill from launch, since the free note only names the three clients, but the wording isn't explicit and we'd check the first invoice.

Is it worth switching to? For batch retrieval at modest volume, probably yes: the per-call price is hard to argue with and the operator set covers what most pipelines need. For anything bursty or multi-tenant, the account-wide cap decides it before the price does. We track model prices and dated cutoffs like this one in the retirements and price cliffs calendar.

FAQ

Sources: DeepSeek, Search & Fetch guide; DeepSeek, API change log (17 October 2026); Anthropic, Web search tool pricing; PPC Land, DeepSeek prices web search for AI agents at $0.75 per 1,000 calls.