DeepSeek search and fetch – $0.75 per 1000 successful requests | AI Paper Digest

TL;DR Highlight

DeepSeek launched web Search and Fetch APIs at $0.75 per 1,000 successful requests

Who Should Read

Backend developers adding live web search and page fetching to agents or RAG who are comparing search API costs

Core Mechanics

Search and Fetch are separate APIs
- The Search API finds web pages for a query and returns structured results.
- The Fetch API takes one URL and returns its content as Markdown.
- Both are JSON APIs using the same bearer auth as the DeepSeek API.
Search returns model-friendly fields
- snippet is 100-200 chars for UI display, and summary is 600-800 chars meant for a model to read.
- Set need_content to true to get page content; with the default false, content is an empty string.
- rerank_score runs 0 to 3, and 1.5 or above counts as high quality.
Query operators pass through intact
- You can use site:, -site:, lang:, after:, before: and exact phrases in the query.
- A query is 1-128 characters, and count is 1-50 (default 10).
Billing is per successful request
- Search and Fetch are each $0.75 per 1,000 successful requests, not billed by tokens.
- Only HTTP 200 with a result is billed; failures, empty results and timeouts are free.

Community Reactions

Several comments compared prices
- Commenters cited $5 for Brave Search API, $12 for Kagi API and $4 for Exa.
- Another comment pointed to an Artificial Analysis table of search API costs.
A commenter said DIY tools are far cheaper
- One built search and fetch MCP tools on third-party scrapers a year ago and found them 15-20x cheaper than OpenAI and Anthropic built-in tools.
- Another said free tools like searxng-ai-kit are more cost effective for personal use.
A benchmark report found quality mediocre
- One commenter ran DeepSeek on their own web-fetch bench and found search and fetch not very good, though cheap.
- They suggested free Jina or paid Parallel, Firecrawl, Tavily and Exa instead.
Questions came up about fetch output and anti-bot
- One user fetched a page with hyperlinks and got only text back.
- Another asked how pages behind Cloudflare or Anubis are handled.

How to Apply

Call Search once with curl
- Today, send a query with count 5 and need_content true using your API key and check summary and rerank_score.
- Run 3 of your agent's real questions and count how many results score 1.5 or above.
Switch your agent search tool to operator queries
- Restrict trusted domains with site: and filter stale pages with after:.
- Feed the model summary first instead of full content to save tokens.
Add retry and 402 handling to your wrapper
- Retry 408, 429 and 5xx with exponential backoff.
- Don't retry 402; alert instead, and block other 4xx until the request is fixed.
Compare against your current provider on the same queries
- Run the same 20 questions through your existing search and fetch provider.
- Tabulate success rate, link preservation and cost per 1,000 requests.

Try It

copy & run

# Paste into a terminal (pip install requests, needs the DEEPSEEK_API_KEY env var)

import os, requests

H = {"Authorization": "Bearer " + os.environ["DEEPSEEK_API_KEY"]}

r = requests.post(
    "https://api.deepseek.com/v1/tools/search",
    headers=H,
    json={"query": "{{question}} lang:en after:2025", "count": 5, "need_content": False},
)
r.raise_for_status()
for x in r.json()["results"]:
    print(x["rerank_score"], x["title"], x["url"])
    print(x["summary"])

# Next: fetch the first result URL and check whether links are preserved
url = r.json()["results"][0]["url"]
p = requests.post("https://api.deepseek.com/v1/tools/fetch", headers=H, json={"url": url})
p.raise_for_status()
print(p.json()["content"][:500])

Terminology

Fetch APIAn API that opens a URL for you and returns the page body as Markdown. Think of it as an errand runner reading pages in place of a browser.

rerank_scoreA 0-3 score combining how relevant, authoritative and fresh a search result is. It's like a report-card grade for each result.

QPSA cap on requests per second. It's like limiting how many people can stand at the counter at once.

Query operatorsSyntax like site: appended to a search to narrow scope. They work like a WHERE clause in SQL.

MCPA protocol that lets agents call external tools in a uniform way. Think of it as a USB port for tools.

DeepSeek search and fetch – $0.75 per 1000 successful requests