Web Search API for AI Agents: What DeepSeek's $0.75 Price Changes

DeepSeek opened its in-house search engine to outside developers on October 17, pricing the new Search and Fetch endpoints at $0.75 per 1,000 successful requests each (PPC Land). If your AI agents rely on a web search API to look things up, that's a sharp drop. The same volume costs $5 on Brave and $8 on Tavily's pay-as-you-go basic tier (DEV Community).

Cheap search is welcome. But picking a search provider for a production agent was never only about price. Result quality, rate limits and where your queries end up matter just as much, and they're usually what bites you three months after launch. Here's how we'd think it through.

TL;DR#

• DeepSeek's Search and Fetch APIs cost $0.75 per 1,000 successful calls, roughly 5 to 10 times cheaper than most alternatives for agent workloads.
• The catches are a 5 requests-per-second cap per account, no public quality benchmarks yet, and servers located in China.
• Put search behind your own tool interface, test providers on your real queries, and route by data sensitivity instead of betting everything on one vendor.

What DeepSeek actually launched#

There are two JSON endpoints on DeepSeek's existing API, both using your normal API key (PPC Land):

• Search takes a query of up to 128 characters and returns 1 to 50 results. Each result carries a short snippet for display, a longer summary written for models, optional full page content, a rerank_score from 0 to 3 and an authority score from 0 to 4.
• Fetch takes a URL and returns the page as Markdown, which is the format language models handle best.

Billing only counts requests that return HTTP 200 with an actual result. Empty results, timeouts and failures are free, and asking for 50 results costs the same as asking for 10.

There's a temporary sweetener too. Calls routed through DeepSeek Harness, Claude Code or Codex are free until 16:00 UTC on October 31, after which normal billing starts. DeepSeek didn't describe a free tier for direct API calls.

RuntimeWire points out that this is the search system behind DeepSeek's own app, now offered to agents and retrieval apps. So this is a production system, not a side project bolted on for launch day.

Why the price matters for AI agents#

A chatbot might search once per question. An agent researching a supplier, checking a competitor's pricing page or verifying an address will often search several times and then open a handful of pages. Costs multiply per task, not per user.

Alan West ran the numbers for 10,000 agent runs at 5 searches and 3 fetches each (DEV Community):

At a few thousand runs a month, $60 versus $450 barely registers. At a million runs, it's the gap between a rounding error and a line item finance asks about.

The bigger effect is behavioural. When search is cheap, teams stop rationing it and let agents double-check more often, which helps as long as the results are good.

The trade-offs behind a cheap web search API#

Rate limits#

Each endpoint is capped at 5 requests per second per account, and extra API keys don't raise it. Fetch also has a per-domain limit DeepSeek hasn't published. A nightly batch job will never notice. A customer-facing agent with 50 people using it at once will hit HTTP 429 errors fast.

Quality is unproven#

Nobody has published a proper benchmark of DeepSeek's index yet. One commenter's informal tests ranked Exa and Tavily highest. This matters more than it sounds: weak results mean more retries and more tokens spent reading irrelevant pages, which can wipe out the savings.

Data residency#

DeepSeek's privacy policy says its servers are in the People's Republic of China, and every query and fetched URL goes there. Regulators have already pushed back: Italy's data protection authority banned DeepSeek's consumer app in January 2025 (PPC Land). By contrast, Brave offers zero data retention on its Enterprise plan.

Fetch behaviour#

The docs don't say which user agent the fetcher uses or whether it honours robots.txt. If your agent fetches pages on behalf of customers, you'll want to know.

What it means for your business#

A more useful question than "which provider is best?" is "which provider fits this particular workflow?"

Some agent work involves only public information. Enriching a lead list with company details from public websites, monitoring competitor pricing pages or pulling product specs for a catalogue are all low-risk. Queries reveal little, volumes are high and nobody is waiting on a reply. That's a good place to try the cheapest option.

Other work is sensitive. A support agent that turns a customer's message into a search query is sending that customer's words to the search provider. A compliance agent researching a specific deal may reveal what you're working on. For those, provider location, retention terms and contracts matter more than a few hundred dollars a month.

Most companies have both kinds of workflow. So the practical move is to stop treating search as a single vendor choice and start treating it as a routing decision.

How to pick (and switch) a web search API: a checklist#

Here's the setup we recommend before you commit to any provider:

agent ──> search_tool (your interface)
              │
              ├─ public batch jobs  ──> low-cost provider
              ├─ customer data      ──> zero-retention provider
              └─ fallback on 429    ──> secondary provider
              │
              └─ log: query class, provider, cost, latency

1. Wrap search behind one interface. Your agent calls search_tool, never a vendor SDK directly. Swapping providers then takes a config change, not a rewrite.
2. Build a test set of 50 real queries from your actual workflows, with what a good answer looks like for each.
3. Measure the whole run, beyond the search bill. Track answer quality, tokens used and retries per task for each provider.
4. Classify every workflow by data sensitivity and route sensitive ones to providers with acceptable residency and retention terms.
5. Handle rate limits properly. Queue requests, back off on 429 and fail over to a second provider instead of failing the task.
6. Filter weak results. Use scores like DeepSeek's rerank_score to drop low-relevance hits before they reach the model.
7. Log cost per run so you can see when a cheaper provider is quietly costing more in tokens.

How MagicMakers Lab approaches this#

When we build agents through our AI Integration & Automation and MCP Development work, external tools like search sit behind a small, provider-neutral layer, often an MCP server that exposes one search tool with routing, logging and approval rules inside it. That keeps the agent simple and the vendor swappable. It's the same thinking behind our data pipeline work for GeoVerdant, where analysis that took weeks now takes minutes because the data sources are wired in cleanly rather than hard-coded.

Key takeaways#

• DeepSeek's $0.75 per 1,000 price makes live web access cheap enough that agents no longer need to ration searches.
• The 5 requests-per-second cap suits batch and internal jobs far better than busy customer-facing agents.
• Search quality hasn't been benchmarked, and poor results can cost more in tokens than they save.
• Queries go to servers in China, so keep customer data and sensitive research on providers with clear retention terms.
• An abstraction layer with per-workflow routing lets you take the savings without locking yourself in.

FAQ#

What is a web search API?#

A web search API lets software query a search engine and get results back as structured data, such as URLs, titles, snippets and page text, instead of a web page built for people. AI agents use one to look up current information, check facts and read source pages before answering or taking an action.

How much does a web search API cost for AI agents?#

Prices vary a lot. DeepSeek charges $0.75 per 1,000 successful requests, Brave charges $5 and Tavily's basic pay-as-you-go tier is $8. Because agents often search and fetch several times per task, the real cost depends on calls per run, so estimate from your workflow rather than from the headline price.

Is the DeepSeek Search API safe for business use?#

It depends on the data. DeepSeek's servers are in China, so every query and URL goes there. For public, low-sensitivity work like lead enrichment or price monitoring, many teams will find that acceptable. For anything involving customer messages or confidential research, check with your compliance team first and consider a provider with zero-retention terms.

Can I switch web search providers without rebuilding my agent?#

Yes, if search sits behind your own tool interface. The agent calls one function, and a small routing layer decides which provider handles each request. Switching or adding a provider then means changing configuration and rerunning your test queries, not rewriting prompts or agent logic.

If your team is weighing which agent workflows can safely use cheaper tools, and which need tighter controls, we're happy to look at your setup with you and map it out. Book a free audit.

Sources#

• DeepSeek prices web search for AI agents at $0.75 per 1,000 calls (PPC Land)
• DeepSeek Search API: $0.75 per 1,000 Calls vs Tavily's $8. Here's the Catch (DEV Community)
• DeepSeek puts web search on the API meter at $0.75 per 1,000 calls (RuntimeWire)