DeepSeek Search API Pricing: Cheap, but Is It Good?

DeepSeek started charging for web search and page fetches made through its agent integrations at midnight Beijing time on Nov 1, 2026, at $0.75 per 1,000 successful calls. That is several times cheaper than the list price of Exa, Brave or Perplexity. The catch, according to independent benchmarks from Artificial Analysis, is that DeepSeek's results are less accurate on hard multi-step research and make the answer model work harder, which eats much of the saving.

What changed on Nov 1

DeepSeek opened its Search and Fetch endpoints to all developers on Oct 17, 2026, alongside DeepSeek V4.1-Flash-1017. The same tools also run server-side when you use DeepSeek models through its agent integrations, including DeepSeek Harness, Claude Code and Codex CLI. Until the end of October those server-side calls were free. From 00:00 Beijing time on Nov 1 (16:00 UTC on Oct 31), they are billed at the same rate as direct API calls.

If you point Claude Code or Codex at DeepSeek's API and let the agent browse, you now pay for every successful search and fetch on top of tokens. For most solo developers that will be cents a day. For teams running agents in CI or on schedules, it is a new line item worth measuring.

How the DeepSeek Search and Fetch API works

Both endpoints live under https://api.deepseek.com/v1/tools/ and use your normal DeepSeek API key. The details below come from DeepSeek's Search & Fetch guide.

• Search (/v1/tools/search) takes a query of 1 to 128 characters and returns up to 50 results (10 by default). It supports site:, -site:, exact "phrase" matching, lang:, after: and before:.
• Each result includes a title, URL, snippet, a 600 to 800 character summary, the date, a rerank score from 0 to 3 (DeepSeek treats 1.5 or higher as high quality) and an authority score from 0 to 4. Setting need_content adds the full page text.
• Fetch (/v1/tools/fetch) takes one URL of up to 8,192 characters and returns the page title and its content as Markdown.
• Rate limits are 100 queries per second for Search and 25 for Fetch, per account across all keys, and Fetch adds a per-domain limit.
• Errors are cheap. A 402 means your balance is too low, a 404 from Fetch means no usable content, and DeepSeek recommends retrying 408, 429 and 5xx responses.

A minimal call looks like this. This is an illustrative example built from the documented parameters, with a placeholder key.

curl https://api.deepseek.com/v1/tools/search \
-H "Authorization: Bearer $DEEPSEEK_API_KEY" \
-H "Content-Type: application/json" \
-d '{"query": "site:docs.python.org asyncio TaskGroup", "count": 5}'

We have written up the endpoints, limits and pricing on a new DeepSeek Search & Fetch API page.

DeepSeek search API pricing vs Exa, Brave and Perplexity

DeepSeek bills only requests that return HTTP 200 with non-empty results or page content. Failures, empty results and timeouts are free, and asking for 50 results costs the same as asking for 10. Here is how the list prices compare. Prices as of Nov 2, 2026, from each vendor's official pricing page.

• DeepSeek Search: $0.75 per 1,000 successful requests. DeepSeek Fetch: $0.75 per 1,000. Empty results are not billed (DeepSeek docs).
• Exa: $4 per 1,000 searches on lite or instant, $7 per 1,000 on auto or fast, $1 per 1,000 extra results above 10, and $1 per 1,000 pages for contents. Exa includes $10 of free credit a month (Exa pricing).
• Brave Search API: $5 per 1,000 requests with up to 50 queries per second, on prepaid credits (Brave Search API).
• Perplexity Search API: $5 per 1,000 requests, or $1 per 1,000 on Fast Search. Perplexity bills a successful request even when it returns zero results (Perplexity pricing).

On list price alone, DeepSeek is about 5 to 9 times cheaper per search than the standard tiers above. But per-search price is the wrong number to optimise, because an agent does not stop after one search. It keeps searching and reading until it can answer, and every result it reads is billed again as model tokens.

What the Artificial Analysis Search Index says

Artificial Analysis runs every search API through the same open-source agent harness, Stirrup, with the same answer model, GPT-5.6 Luna at medium effort. Its Search Index is the average of three tests: DeepSearchQA (multi-step research, scored as F1), BrowseComp (hard multi-hop browsing) and AA-Omniscience (factual recall, which we covered in our hallucination benchmark breakdown). Because the model and harness are fixed, the scores compare the search tools rather than the models.

Artificial Analysis added DeepSeek Search on Oct 22, 2026. It scored 69 on the index, which it ranked 17th of 27 search products at the time and the lowest-priced paid search API on its board.

The figures below are from the Artificial Analysis Search API leaderboard as read on Nov 2, 2026 (latest data dated Oct 19, 2026, with DeepSeek added Oct 22). Each line shows index score, then search cost and model cost per 1,000 tasks.

• Exa (auto): index 80. DeepSearchQA 86, BrowseComp 83, Omniscience 73. $53.17 search plus $12.43 model, about $65.60 total.
• Perplexity (medium): index 80. $62.30 search plus $8.04 model, about $70.34 total.
• Exa (lite): index 79. $31.20 search plus $13.18 model, about $44.38 total.
• Octen (highlights): index 77. DeepSearchQA 80, BrowseComp 86, Omniscience 66. $9.07 search plus $14.92 model, about $23.99 total, and the fastest at 16.0 seconds per task.
• OpenAI Web Search (medium): index 74. $40.19 search plus $9.40 model, about $49.59 total.
• Firecrawl: index 73. $30.48 search plus $12.09 model, about $42.57 total.
• DeepSeek Search (summary): index 69. DeepSearchQA 75, BrowseComp 63, Omniscience 70. $8.47 search plus $23.24 model, about $31.71 total, at 41.2 seconds per task.
• No search (model only): index 33, with $2.92 model cost.

Totals are our sums of the two cost columns Artificial Analysis publishes. These numbers apply to this harness and this answer model. Your own costs will differ with a different model, prompt or agent, so treat them as a relative ranking rather than a quote.

Cheap search, expensive tokens

Two things stand out. First, DeepSeek is competitive on factual lookups but weak on hard browsing. Its 70% on AA-Omniscience is three points behind the leaders, while its 63% on BrowseComp is the lowest of the 12 products shown on the leaderboard, where the other 11 score between 74% and 87%. If your agent mostly checks docs, prices or release dates, that gap may not matter. If it chases multi-hop research questions, it will.

Second, DeepSeek has the highest model cost per task of the products listed. Artificial Analysis puts the total at about $0.032 per task, with only 27% of that spent on search. It attributes the rest to DeepSeek's larger results, which give the answer model more to read. The same effect shows up in speed. A DeepSeek search returns in about 0.5 seconds on average, yet a full task takes about 40 seconds because 34 seconds of it is model time, the most on the board.

So the cheapest search is not the cheapest task. Octen costs less per 1,000 tasks than DeepSeek in this benchmark and scores eight points higher, and Exa lite scores ten points higher for about $13 more per 1,000 tasks. If your answer model is pricier than GPT-5.6 Luna, the token side of the bill grows and DeepSeek's advantage shrinks further. If you route through a cheaper model or a gateway with caching, as we discussed in our AI gateways comparison, the maths moves the other way.

What to do if you run Claude Code or Codex on DeepSeek

• Check whether your agent browses at all. Many coding sessions never search the web. If yours do, the new charge shows up in your DeepSeek usage and billing alongside token spend.
• Do a rough budget. As an illustration only, an agent that makes 20 searches and 10 fetches per task across 1,000 tasks a month makes 30,000 billable calls, which is $22.50 at $0.75 per 1,000. Token costs for reading those results are usually larger.
• Use operators to cut noise. site: and lang: filters return fewer, more relevant results, which means fewer tokens for the model to read.
• Trim what you pass on. Use the summary or snippet first and fetch full pages only when the rerank score is high. That attacks the model-token cost, which is the bigger half of the bill.
• Pick by task. For quick factual checks inside a coding agent, DeepSeek is a reasonable low-cost default. For deep research agents where accuracy on hard questions matters, test a higher-scoring API on your own prompts first. Our guide to open-weight coding models covers where DeepSeek's models fit on the model side.

Verdict

DeepSeek's Search and Fetch API is the cheapest paid search option on the Artificial Analysis board, and the Nov 1 billing change is small for most developers. But its 69 index score and weak BrowseComp result mean it is not the best value for hard research, because the answer model ends up doing more work. Use it for cheap factual lookups inside coding agents, measure your real token spend, and test Octen or Exa lite if your agent needs to dig deeper.

Sources

• DeepSeek API Docs: Search & Fetch guide (endpoints, pricing, billing rules, rate limits), accessed Nov 2, 2026
• DeepSeek on X: Search & Fetch APIs open to everyone, Oct 17, 2026
• DeepSeek API Docs: Using the Anthropic API, accessed Nov 2, 2026
• Artificial Analysis: Search API leaderboard and Search Index, accessed Nov 2, 2026
• Artificial Analysis on X: DeepSeek Search debuts at 69, Oct 22, 2026
• Artificial Analysis on X: DeepSeek Search cost per task, Oct 22, 2026
• Exa pricing, accessed Nov 2, 2026
• Brave Search API pricing, accessed Nov 2, 2026
• Perplexity API pricing, accessed Nov 2, 2026
• Neomanex: DeepSeek Search & Fetch API billing starts Nov 1, accessed Nov 2, 2026