DeepSeek Opens Search and Fetch APIs at $0.75 per 1,000 Calls, 13x Cheaper Than OpenAI

TL;DR

DeepSeek opened two plain JSON endpoints on October 17, a web Search API and a page Fetch API, each priced at $0.75 per 1,000 successful requests. OpenAI and Anthropic both charge $10 per 1,000 for their hosted web search tools, Brave and Perplexity charge $5, and Exa charges $4 to $15 depending on mode. You pay only when a call returns HTTP 200 with a result, the number of results does not change the price, and each endpoint is capped at five requests a second. Searches made through the server-side tool in DeepSeek Harness, Claude Code or Codex are free until billing starts at 00:00 Beijing time on November 1.

What shipped

The two endpoints live at POST https://api.deepseek.com/v1/tools/search and POST https://api.deepseek.com/v1/tools/fetch, and they take the same bearer key as the chat API. They arrived in the same October 17 release note as DeepSeek-V4.1-Flash-1017, the model refresh that took the Hacker News front page. The search half carries a price nobody else in the market matches.

DeepSeek says both run on its in-house search engine, which it claims already serves hundreds of millions of requests a day. That number describes volume. The docs publish no relevance benchmark.

A search request takes a query of 1 to 128 characters, a count from 1 to 50 (default 10), an optional ori_question carrying the user's original wording, an IP address in localization to localize results, and a need_content flag that is off by default. The query string accepts site:, -site:, quoted phrases, negated terms, lang:, after: and before:.

Each result comes back with a rerank_score from 0 to 3 (DeepSeek calls 1.5 and above high quality), an authority score from 0 to 4 (-1 means unknown), a snippet of roughly 100 to 200 characters, a summary of roughly 600 to 800 characters, the page date, and the full page text only if you asked for it. Fetch takes one HTTP or HTTPS URL of up to 8,192 characters and returns a title plus the page as Markdown. Both endpoints are read-only.

The price next to everyone else's

Hosted web search for agents mostly runs $4 to $10 per 1,000 calls. OpenAI's pricing page lists web search at $10.00 per 1,000 calls for all models, plus the search content tokens billed at model rates. Anthropic's web search tool is $10 per 1,000 searches plus tokens, with each search counted once regardless of how many results it returns; its web fetch tool has no per-call charge. Brave's Search API is $5 per 1,000 requests with $5 of free credit a month and a 50 requests-per-second limit. Perplexity's Search API is $5 per 1,000 for standard search and $1 for its fast tier. Exa charges $4 per 1,000 for instant search, $7 for fast and auto, $12 for deep and $15 for deep-reasoning, plus $1 per 1,000 pages of contents. SerpAPI's Production plan works out to $10 per 1,000.

At one million searches a month, that is $750 on DeepSeek against $10,000 on OpenAI or Anthropic and $5,000 on Brave. For a search-then-read loop, the pair costs $1.50 per 1,000 on DeepSeek, about $8 on Exa (auto search plus contents) and $10 on Anthropic, where the fetch is free but the search is not. Perplexity's $1 fast tier is the only published price within shouting distance, and it is still a third more.

How the meter runs

DeepSeek bills a request only when it returns HTTP 200 and produces a result: a non-empty results array for search, page content for fetch. Failed calls, empty result sets and timeouts cost nothing. The count parameter does not change the price, so 50 results cost the same $0.00075 as 10, though with need_content on you will pay in model tokens for whatever you feed back to the model. The limits are five requests a second per endpoint per account, with an extra per-domain cap on fetch, and anything above that gets a 429. Fetch can also return a 403 when DeepSeek's security policy rejects the target, or a 404 when a page yields no usable content.

The pricing page notes that DeepSeek may adjust prices, which is the usual disclaimer, and the token side of the same page is already a dual-rate schedule: deepseek-flash costs $0.15 per million input tokens off-peak and $0.30 at peak, with output at $0.60 and $1.20. The search and fetch lines carry a single flat rate.

Free in the harnesses until November 1

If you already point Claude Code or Codex at DeepSeek, the search tool is a server-side call: the model asks for a search, the same API that runs the model runs the search and hands the results back, and your terminal only sees the answer. DeepSeek's Claude Code guide says the API natively supports Claude Code's Web Search feature through the Anthropic-compatible endpoint at https://api.deepseek.com/anthropic, and warns that each search triggers extra LLM requests you pay for in tokens. The Codex guide only enables built-in web search on deepseek-v4-pro, with web_search set to "live" in config.toml; DeepSeek's own setup script leaves it off.

Those server-side calls, plus the ones DeepSeek Harness makes, are free for DeepSeek API users right now. Per-request billing starts at 00:00 Beijing time on November 1, 2026, which is 16:00 UTC on October 31. DeepSeek's idea of a launch promotion is to charge nothing and publish the exact minute that stops.

What the docs do not tell you

There is no published relevance benchmark, and the two scores you get per result are DeepSeek's own scales, so you will need to calibrate them against your queries. The Hacker News thread, at 531 points and 254 comments when we checked, asked the practical questions the docs skip: how fetch copes with Cloudflare and Anubis challenges, whether results are censored, and whether page authors get anything for being fetched. None has an official answer. One commenter reported that a fetched page came back as text with the hyperlinks stripped, which matters if your agent follows links; the fetch docs describe Markdown output and say nothing about link handling.

Five requests a second is 18,000 searches an hour, plenty for a research agent and too little for a consumer product with concurrent users, a point the thread argued both ways. Every query and every fetched URL also goes to a Chinese provider's servers, which raises a data-handling question for anything touching client data.

What to do with it

If your agent already calls a search tool, this is the new floor, and a swap is a thin adapter because both endpoints are raw JSON with no SDK required. Keep need_content off, prune on rerank_score before you fetch, and fetch only what the model will read; the price makes over-fetching cheap, the token bill does not. Run your own side-by-side against Brave or Exa before you commit, which is affordable in a way it was not last week: a 10,000-query eval costs $7.50. And if you run Claude Code or Codex on DeepSeek, you have 12 days of free search to find out whether the index is good enough to keep.

Key Takeaways

• DeepSeek's Search and Fetch APIs went live October 17 at $0.75 per 1,000 successful requests each, as raw JSON endpoints under the normal API key.
• Most competitors charge 5x to 13x more: OpenAI and Anthropic $10 per 1,000 searches, Brave and Perplexity $5, Exa $4 to $15 by mode, SerpAPI about $10 on its Production plan. Perplexity's $1 fast tier is the exception.
• You pay only for success: HTTP 200 with a result; empty, failed and timed-out calls are free, and 50 results cost the same as 10.
• Limits: five requests a second per endpoint per account, a per-domain cap on fetch, queries capped at 128 characters.
• Free in Claude Code, Codex and DeepSeek Harness until 00:00 Beijing time on November 1; Codex only gets web search on deepseek-v4-pro.
• Unknowns: no relevance benchmark, no word on censorship or link handling in fetch, and your queries go to a Chinese provider.

Sources: DeepSeek API docs: Search & Fetch, DeepSeek API docs: Models & Pricing, DeepSeek release note, October 17, 2026, DeepSeek docs: Claude Code integration, DeepSeek docs: Codex integration, DeepSeek-V4.1-Flash-1017 on Hugging Face, OpenAI API pricing, Anthropic web search tool docs, Anthropic web fetch tool docs, Brave Search API pricing, Perplexity API pricing, Exa pricing, SerpAPI pricing, Hacker News discussion