DeepSeek opened its in-house search engine to developers on October 17 at $0.75 per 1,000 successful requests, for both a Search endpoint and a Fetch endpoint that turns a page into Markdown. DeepSeek’s own documentation says only requests that return results are billed: empty result sets, failures and timeouts are free. Against the $10 per 1,000 that Anthropic and OpenAI charge for their built-in search tools, that looks like a 92% discount.
It mostly isn’t, and the reason is the most useful thing to understand about web search for AI agents right now. The per-call fee is the smallest line on the bill. What decides the cost is how much text each search drags into the model’s context, and how many times that text gets re-sent before the task finishes.
The listed prices, and what each one leaves out
Here is where the main options sit. The third-party rows are list prices compiled by Digital Applied in early October, with no enterprise discounts; check each vendor’s page before you commit, because these move.
| Route | Listed price per 1,000 | What you also pay |
|---|---|---|
| DeepSeek Search API | $0.75 (successful only) | Fetch is a separate $0.75 per 1,000 |
| Brave Search API | $5 | Nothing; $5 of free credit monthly |
| Perplexity Search API | $5 | No token costs listed |
| Exa (direct) | $7 | $1 per 1,000 for each result above 10 |
| Tavily (basic / advanced) | $8 / $16 | Nothing |
| OpenAI web_search tool | $10 | Search content tokens at model rates |
| Anthropic web search tool | $10 | Results billed as input tokens, every turn |
| Anthropic web fetch tool | $0 | Fetched page billed as input tokens |
| Google grounding, Gemini 3.x | $14 | Model tokens; one prompt can run several queries |
Brave’s figures come from its API page, which also lists a 50 queries per second limit on the Search plan. OpenAI’s pricing page states the web search tool costs “$10.00 / 1k calls + Search content tokens billed at model rates,” and for two older mini models it bills search content as “a fixed block of 8,000 input tokens per call,” according to OpenAI’s pricing documentation.
The right column is where the money is. Two of the cheapest-looking rows (Anthropic’s free fetch and DeepSeek’s $0.75 search) can be among the most expensive in practice, depending on what you do with the output.
Where the real cost hides: tokens that ride along
Anthropic’s documentation is unusually direct about this. Search costs “$10 per 1,000 searches, plus standard token costs for search-generated content,” and results “are counted as input tokens, in search iterations executed during a single turn and in subsequent conversation turns,” per the web search tool reference. The sample response on that same page shows 6,039 input tokens for a single search.
Run that through Claude Sonnet 5.5, which lists at $2 per million input tokens and $0.10 for cache hits on Anthropic’s pricing page. The arithmetic below is ours, and it is illustrative, not a measured workload:
- One search: $0.01 in fees, about $0.012 in tokens. The tokens already cost more than the search.
- A research turn with 10 searches pulls in roughly 60,000 tokens. If the agent then takes five more turns, those tokens ride along each time. Uncached, that is about 360,000 input tokens, or $0.72, against $0.10 in search fees.
- With prompt caching working, the five follow-up sends drop to about $0.03 in total. The difference between a cached and an uncached agent is bigger than the difference between any two search vendors in the table.
Fetch is worse. Anthropic charges nothing per fetch, but a full documentation page or a filing can run to tens of thousands of tokens, and it stays in context. That is why the fetch tool ships a max_content_tokens parameter, and why the docs say to set it “to protect against inadvertently fetching large content that would consume excessive tokens.”
DeepSeek’s design leans the other way. Each result carries a summary of roughly 600 to 800 characters written for models, and full page content only arrives if you set need_content to true. Ten summaries come to roughly 7,000 characters, which at the usual four characters per token is something under 2,000 tokens. That compactness, not the $0.75, is the real argument for it.
The levers that actually move the bill
If your agent uses a model-native tool, the per-call fee is fixed and the controls are all about volume and payload:
- Cap searches per request. Anthropic’s
max_usesis a hard ceiling. Its docs say simple factual queries “typically use 1–3 searches” while comparative research “can use 10 or more.” Set the cap from the task, not the default. - Filter before it lands in context. Anthropic’s newer
web_search_20260209version lets Claude write code that filters results before they reach the context window, which the docs say “reduces token use on search-heavy requests.” Theweb_search_20260318version addsresponse_inclusion: "excluded"to stop raw results being echoed back to your client. - Cap fetches. Set
max_content_tokenson every fetch tool. An uncapped fetch is the single easiest way to blow a budget. - Restrict domains.
allowed_domainsnarrows both cost and the junk your agent reads. It also helps with a problem cheap search does not touch: agents citing working links for claims those pages never made, which we covered in our look at how deep research agents get citations wrong. - Make sure caching is on. Search results are exactly the kind of large, repeated prefix that prompt caching exists for. Check the cache hit rate on a search-heavy run before you blame the vendor.
If you run your own retrieval loop against an external API, pick by payload shape first and price second. An API that returns only titles and snippets is cheap per call but forces a second fetch for anything substantive, so you pay twice. One that returns model-ready summaries can be the cheaper route even at a higher listed price.
Three things to check before switching to the cheap option
The free period ends November 1. DeepSeek’s docs say server-side tool calls made through DeepSeek Harness, Claude Code or Codex currently hit Search and Fetch “at no charge,” with billing starting at 00:00 Beijing time on November 1 (16:00 UTC on October 31). If you point a coding agent at DeepSeek’s endpoint, its web searches are now DeepSeek searches. DigitalToday’s report makes the practical warning: coding tools search and fetch on their own, so calls pile up even when nobody asked for a search. Meter it now, while it is free, so the first invoice is not a surprise.
Queries are data. A search query tells the provider what your agent is working on: the acquisition target, the vulnerability, the competitor. DeepSeek’s search documentation says nothing about data retention or server region. Brave sells zero data retention on its Enterprise plan and calls it architectural rather than contractual; Anthropic documents which web search configurations are eligible for zero data retention. If your agent touches anything sensitive, the retention terms belong in the evaluation, not after it.
Some exits are closing. Microsoft retired its public Bing Search APIs on August 11, 2025, as Digital Applied notes, and its replacement only works inside Azure. Google’s Custom Search JSON API stops working on January 1, 2027, according to heise, with full-web access only through a custom-quoted enterprise offer. Any agent pipeline still built on either needs a replacement this quarter, which is a good moment to price on tokens instead of calls.
The cheapest search API is the one whose results your model reads the fewest times. Measure cost per completed task with tracing on, the same discipline we laid out for agent observability, and the vendor choice usually becomes obvious. Most of the time it turns out to matter less than the cap you forgot to set.