DeepSeek prices web search at $0.75 per 1,000 calls, 13.3x under OpenAI and Anthropic. On a search agent that line is 90% of a Claude Haiku 5.5 bill, and you can buy it without DeepSeek's model.

DeepSeek now sells web search and page fetch as standalone APIs at $0.75 per 1,000 calls each. OpenAI and Anthropic charge $10 per 1,000 searches, so DeepSeek's search line costs about a thirteenth of theirs. On the agent workload we price below, search fees are 90% of the bill on Claude Haiku 5.5 or GPT-6 Luna. Swap in DeepSeek's tools and the month falls from $335 to $72.50, even if you keep Haiku as the model.

The model in the release, V4.1-Flash-1017, kept the token price. Calls from Claude Code, Codex and DeepSeek Harness are free until October 31 at 16:00 UTC.

What DeepSeek actually launched

Two endpoints, both on the regular API key. POST /v1/tools/search takes a query of up to 128 characters and returns between 1 and 50 results, each with a title, URL, snippet and a summary, plus the full page text if you set need_content. POST /v1/tools/fetch takes a URL and gives the page back as Markdown. DeepSeek says both run on its own in-house search engine.

The billing rules are short and friendlier than most. You pay $0.75 per 1,000 requests that return HTTP 200 with something in them. A search with zero results costs nothing, and so does a timeout or a fetch that comes back empty. Asking for 50 results costs the same as asking for one.

The catch is throughput. Each endpoint is capped at 5 requests a second per account, whichever key you use, and Fetch has a per-domain limit on top of that. That's about 13 million searches a month if you run flat out around the clock. It's plenty for one product. For a crawler or a busy multi-tenant agent platform it may not be.

Every search price we could find

Search is priced per call nearly everywhere, which makes it easy to compare. Here are the model providers' built-in tools next to the standalone search APIs people already wire into agents:

Each provider's pricing page, read October 18, 2026. Tavily converted at $0.008 per pay-as-you-go credit. Serper credits are prepaid packs. Moonshot's older built-in $web_search tool ($5 per 1,000) is left out because Moonshot expects to deprecate it on October 20.

So DeepSeek isn't the cheapest search you can buy. A Serper pack bought in bulk beats it. Among companies that also sell you a model, only Perplexity's fast mode at $1 comes close. xAI, OpenAI, Anthropic and Google charge between 6.7 and 18.7 times as much, and DeepSeek asks for no prepaid volume to get there. Serper only matches $0.75 once you buy 500,000 credits up front.

Fetch is less one-sided. Anthropic's web fetch tool and Gemini's URL context tool charge nothing per call, only the tokens the page adds to your prompt. DeepSeek charges $0.75 per 1,000, and Perplexity's Agent API fetch_url undercuts it at $0.50. Exa sells page contents at $1 per 1,000, Moonshot's Fetch is $2 and Tavily's extract works out to $1.60. OpenAI doesn't list a fetch tool at all.

A month of a research agent

Per-call prices only mean something next to the tokens. So take a plain research assistant: 10,000 questions a month, each one running three searches and two page fetches, reading about 30,000 tokens of results and writing a 1,000-token answer. No caching, since the pages change every time. Every request stays under Haiku's 100K-token price step. The bars split each bill into tool fees (solid) and tokens (faded):

Our arithmetic. 30,000 searches, 20,000 fetches, 300M input and 10M output tokens. DeepSeek Flash at $0.15/$0.60 off-peak and $0.30/$1.20 peak; Haiku 5.5 and GPT-6 Luna at $0.10/$0.50; Gemini 3.8 Flash at $0.75/$3.75 until December 31. Gemini's tool line is 25,000 billable searches at $14 per 1,000 after the free 5,000. Different tokenizers count the same page differently, so the token halves are approximate.

Look at the Haiku and Luna rows. Both models are cheaper per token than DeepSeek Flash: $35 of tokens against $51 off-peak. Then the search line adds $300 and the cheap model ends up 3.8x as expensive. Three searches per question at a cent each costs more than 30,000 tokens of reading at $0.10 per million. Most budget-model comparisons price the tokens and skip that line.

Gemini is the worst of the set here, even with its free 5,000 searches. Its grounding fee is the highest per call, and 3.8 Flash's tokens cost 7.5x Haiku's. When the introductory rate ends on January 1, the token half doubles and this month goes from $612.50 to $875.

The top row is the one we'd actually run. DeepSeek's tools are plain HTTP endpoints, not a server-side feature tied to its model, so you can call them from your own loop and hand the results to Haiku 5.5 or GPT-6 Luna. That's $72.50 a month. You lose the convenience of the provider running the search loop for you. You also send every query to a second company, which some teams won't be allowed to do with a China-based one.

The free window, and what it costs after

If you point Claude Code, Codex or DeepSeek Harness at DeepSeek's API, the agent's built-in web search and fetch calls now go to these endpoints. They're free until 16:00 UTC on October 31. DeepSeek's pricing page gives the same moment as midnight on November 1, Beijing time.

After that, the bill is small. A long coding session that searches or fetches 40 times costs 3 cents in tool fees, against 40 cents for the same 40 searches on Anthropic's tool. For a coding agent the tokens will dwarf either figure. It only becomes a big line when search is the product, as in the month above.

Same weights size, same price, higher scores

The model half of the release is a swap behind an unchanged name. Since October 17, deepseek-flash runs DeepSeek-V4.1-Flash-1017, and the retired deepseek-v4-flash aliases route there too. The rate card didn't move: $0.15 input, $0.003 cached and $0.60 output off-peak, all doubled during the weekday peak windows (01:00 to 04:00 and 06:00 to 10:00 UTC, Chinese public holidays excepted).

DeepSeek's own numbers from the Hugging Face model card. Nobody has rerun them independently yet.

An 11-point jump on Terminal-Bench 4.0 is the one that stands out, even if it's still 9.5 behind Opus 5. DeepSeek also says the new build writes shorter outputs, which would cut cost per task at the same token price. It backed that with a chart of score against cost per task and no figures, so we can't check it yet.

The weights are on Hugging Face under MIT, and the checkpoint is the same size as September's: about 763 billion parameters, which DeepSeek's September card breaks into a 552B backbone plus 196B of what it calls Engram conditional memory, with 8B active per token on input and 16B on output. So this is a retrain of the same architecture, and if you already self-host V4.1 Flash, your hardware bill doesn't change.

Where we land

If your agent searches more than a couple of times per answer, price the search line before you pick the model. At $10 per 1,000, search will cost more than the tokens on any budget model, and DeepSeek's $0.75 is the cheapest search we can find without a prepaid pack. We'd test result quality on your own queries first, because a search engine nobody outside DeepSeek has benchmarked could need a fourth search to match what three Google results give you. That would still cost a fraction of one $10 call.

For the token half of the bill, compare budget models on the pricing page or plug your own volumes into the calculator.

Pages behind the numbers

• DeepSeek: V4.1-Flash-1017 release note - release date, Search and Fetch launch, free window
• DeepSeek: Search & Fetch guide - endpoints, billing rules, rate limits
• DeepSeek: models and pricing - token card, peak hours, tool prices
• Hugging Face: DeepSeek-V4.1-Flash-1017 - benchmarks, weights, license
• OpenAI pricing, Anthropic pricing and Gemini API pricing - built-in search and fetch tools, model token rates
• xAI, Perplexity, Brave, Exa, Tavily and Serper - standalone search prices
• Moonshot: tool pricing - Kimi web search deprecation