Best way seriesNovember 16, 2026
The best way to search the web with an AI agent
Use exa/search at 1¢ for a lookup and one vaaya/onesearch call at 5¢ when the answer matters. Keep deep search for questions one pass cannot answer.
For a quick lookup, have your agent call exa/search (1¢) with a date filter. When the answer matters, make one vaaya/onesearch call (flat 5¢): it can race a second index and read the pages it finds, and every item of evidence it returns names its source. Keep linkup/deep-search (5¢) and parallel/task (10¢ or 30¢) for questions that one pass cannot answer.
What best means for a web search
Four things decide whether a search did its job.
- Freshness. The results are as recent as the question needs.
- What comes back. Links with snippets, page text, or an answer with sources. Snippets mean a second call to read the page.
- Corroboration. A second, independent index agrees with the first.
- Cost per answer. The total across every call the answer took.
The search tools in the catalog
Prices are as listed on 16 November 2026. The catalog has current prices.
| Tool | What it returns | Price per call | When to pick it |
|---|---|---|---|
exa/search |
Semantic search results, with page text inline if you pass contents: { text: true } |
1¢ | The default for finding pages on a topic. Takes date and domain filters. |
brave/search |
Keyword results with snippets, from Brave's own index | 1¢ | An independent second source for a fact that matters |
serper/search |
Google results: organic ranks, knowledge graph, people-also-ask | 0.1¢ | The question is what Google ranks for a query |
perplexity/search |
Results with snippets, each carrying date and last_updated |
0.5¢ | How recent a page is decides the answer |
tavily/search |
Results with relevance scores, plus a synthesized answer if you pass include_answer |
0.8¢ | You are feeding results to a model |
linkup/search |
A cited answer by default, or raw ranked results | 1¢ | You want an answer with sources instead of a list of links |
valyu/search |
Results with full-text excerpts already extracted | 2¢ | You would otherwise search, then fetch the body text |
vaaya/onesearch |
Evidence items with URL, title, snippet, optional full content and the source tool | 5¢ flat | One call instead of a search, a second index and a page read |
linkup/deep-search |
The same output as linkup/search, from iterative multi-query retrieval |
5¢ | A one-pass search came back thin on a multi-hop question |
parallel/task |
An asynchronous research run. Poll parallel/task-status (free) for the result |
10¢ on pro, 30¢ on ultra |
A question that needs a full research report |
vaaya/onesearch-deep |
Asynchronous. Evidence ranked and corroborated across sources | It holds your budgetCents (5 to 500) and charges only what its sources cost |
A hard question the flat call under-covers |
If you leave budgetCents out of vaaya/onesearch-deep, the hold is 10¢ at depth: "standard", 50¢ at "deep" (the default) and $1.50 at "exhaustive".
The default path
Start with semantic search. exa/search is the default for finding pages on a topic. Pass contents: { text: true } to get page text with the results, and set startPublishedDate for anything time-sensitive.
Add a keyword index when a fact matters. brave/search runs on Brave's own index, separate from Google and Bing, so it is an independent check. Compare the two result sets by URL. parallel/search (1¢) is another second source. Use serper/search when the question is what Google ranks.
Ask for a cited answer when you want the answer and its sources in one step. linkup/search returns one by default. For papers, valyu/academic (1¢) searches arXiv and PubMed and returns the paper text.
Go deep only after a one-pass search comes back thin. linkup/deep-search is for multi-hop questions. parallel/task is the larger step, for a question that needs a full research report.
Use vaaya/onesearch when you would otherwise run that chain yourself. Give it a plain-English query and it plans the retrieval. Add timeCritical to race two independent indexes, fidelityRequired to fetch full page content, or facets: ["docs"] to get documentation back as complete markdown. The internal source calls are included in the 5¢, and the call is not charged if every source fails. It also reaches sources the search vendors do not, such as SEC filings and case law.
A worked example
This is arithmetic at listed prices on 16 November 2026. The question: what changed in the newest stable PostgreSQL release, with a link to the release notes.
The chain starts with one search:
{
"service": "exa",
"action": "search",
"params": {
"query": "PostgreSQL newest stable release, what changed",
"numResults": 10,
"contents": { "text": true },
"startPublishedDate": "2026-08-16"
},
"max_cost_cents": 1
}
exa/searchwith a date filter and inline text: 1¢.brave/searchwithfreshness: "pm"as the independent check: 1¢.exa/contentsto read two pages that only Brave returned: 2 URLs at 0.1¢ each, rounded up to the 1¢ minimum.
Total: 3¢ across three calls.
The single call:
{
"service": "vaaya",
"action": "onesearch",
"params": {
"query": "What changed in the newest stable PostgreSQL release?",
"timeCritical": true,
"fidelityRequired": true
},
"max_cost_cents": 5
}
That is 5¢. The chain costs less in cents. OneSearch is one call at a price you know in advance, and the agent does no orchestration. If either comes back thin, linkup/deep-search adds 5¢ and a parallel/task run on pro adds 10¢.
At these prices, 200 single searches cost $2 and 200 OneSearch calls cost $10.
How to run it
Connect your agent once. The setup guides for MCP, the API and the SDK are at /install. Then ask in plain English:
Find out what changed in the newest stable PostgreSQL release. Use sources from the last three months, link the official release notes and one independent source, and spend at most 10 cents.
The agent can ask Vaaya's consult tool to plan the calls, and it gets back the calls to run with their prices. Each call the agent then makes carries max_cost_cents. A call priced above that number is refused before it runs. For a tool charged a fraction of a cent, such as serper/search, Vaaya still checks and holds 1¢ first, so max_cost_cents must be at least 1.
What to check, and where this falls short
- Dates. Check the publish date on every source you use. Without a date filter, a semantic search can return stale pages for a question about the latest version of anything.
- Snippets.
brave/search,serper/searchandperplexity/searchreturn snippets with no page content. Read the page before you quote it. - Agreement. One index can miss a page another has. For a fact you will act on or publish, run a second index and prefer primary sources over aggregators.
- Cited answers. An answer from
linkup/searchis a summary. Open the sources it cites. - Deep runs.
linkup/deep-searchis slower than a one-pass search, andparallel/taskis asynchronous. Neither suits a simple lookup. - Scope. OneSearch answers questions. For people, or for URLs you already hold, use a different tool.
For a recurring digest, the news monitoring guide covers brave/news (1¢) and serper/news (0.1¢). For a longer research run with a spending limit on the loop, see the research run walkthrough.
Why pay per call suits search
Search demand arrives in bursts: nothing for a day, then a few hundred queries when a task lands. Paying per call means the quiet day costs nothing, and one task can use several indexes without you opening an account or holding an API key with any of them. A daily, weekly or monthly limit on the agent's key stops a search loop at a number you chose (see spending controls), and max_cost_cents keeps a 1¢ lookup from turning into a 30¢ research run. Pricing lists the prepaid packs.
Questions
What does a web search cost an AI agent?
Through Vaaya, one search with exa/search, brave/search or linkup/search is 1¢, tavily/search is 0.8¢, perplexity/search is 0.5¢ and serper/search is 0.1¢. One vaaya/onesearch call is a flat 5¢. A parallel/task research run is 10¢ or 30¢. Prices are as listed on 16 November 2026.
Should an agent use keyword search or semantic search?
Start with semantic search (exa/search) to find pages on a topic. Add a keyword index such as brave/search as an independent second source for facts you will act on, and use serper/search when the question is what Google ranks.
When is one OneSearch call better than several searches?
When you would otherwise chain a search, a second index and a page read. vaaya/onesearch does that in one call for a flat 5¢ and names the source of every item it returns. For a single quick lookup, a 1¢ search is enough.
How do I stop a search loop from overspending?
Give the agent its own key with a daily, weekly or monthly limit, and have it send a maximum price with every call. A call priced above that maximum is refused before it runs.