Skip to content

Search API for AI Agents: Web Results with the Passages That Answer

Give your agent web search in one call. POST /v1/search returns each result with the passages of the page that answer the query, in about three seconds, for 0.5 credits a search: $0.50 per 1,000 searches on Max. Pro Search digs deeper into harder questions, for 2 credits.

Answer accuracy and price of search APIs

November 15, 2026

Answer accuracy on FRAMES

↖ Up and to the left: more accurate for less

40%50%60%70%80%$0$2$4$6$8$10$12Firecrawl43.5% · $1.20Tavily basic48.5% · $5.00Tavily advanced53.5% · $10.00Parallel fast63.5% · $1.00Exa68.0% · $7.00Parallel advanced70.0% · $5.00querying.ai Search66.5% · $0.50querying.ai Pro Search68.5% · $2.00

Price per 1,000 searches · lowest public price

Cost of 1,000 correct answers

price ÷ accuracy on 100 company-news questions

  1. querying.ai Search$0.51
  2. Parallel fast$1.00
  3. Firecrawl$1.26
  4. querying.ai Pro Search$2.02
  5. Parallel advanced$5.00
  6. Tavily basic$5.10
  7. Exa$7.07
  8. Tavily advanced$10.20

How it was measured

200 FRAMES questions on November 15, 2026. Each API searched a question once and returned ten results with page text: querying.ai up to 10,000 characters a result, Exa up to 20,000, Tavily, Parallel and Firecrawl their default content. Results that reproduce the benchmark's own questions were removed, then one language model wrote an answer from the rest alone and a second model graded it with the benchmark's grading prompt. Pro Search, Exa and Parallel advanced are within sampling error of one another, and so are Search and Parallel fast. So are Tavily basic and Firecrawl, which was measured the same way on November 17, 2026.

Company news, 100 questions

The public questions of Openbenchmarks' OB-News-Websearch set, asked as its open-source harness asks them: one search per question and ten results, each cut to 1,200 characters. A language model answered from those results alone and the answer was graded against the set's answer key. Every service answered 95 to 100% correctly, so the cost of a correct answer follows the price. Openbenchmarks ranks its own leaderboard on private questions, so these scores are not comparable to it.

SimpleQA, 500 questions

Measured the same way: querying.ai Search 95.8%, querying.ai Pro Search 95.6%, Exa 95.4%, Parallel advanced 96.2%, Parallel fast 95.8%, Tavily advanced 96.8%, Tavily basic 87.4%, Firecrawl 87.0% (first 200 questions). Every service except Tavily basic and Firecrawl scored between 95% and 97%.

Prices on November 15, 2026

Each service's lowest public price for 1,000 searches of up to ten results: querying.ai on the Max plan, Exa search with contents, Tavily on its Growth plan, Parallel at its list price, Firecrawl on its Scale plan billed yearly.

Question sets: FRAMES · SimpleQA · OB-News-Websearch

Endpoint
POST /v1/search
Response time
About 3 seconds, about 4 for Pro Search
Price
0.5 credits a search, 2 for Pro Search
Each result
Address, title, score and the passages that answer

Search features

The features where the services differ, from each service's API reference.

Featurequerying.aiParallelFirecrawlTavilyExa
Date rangeResults published between two dates.YesNoYesYesYes
Deeper searchA deeper mode for questions that need several sources.Yes$2.00 per 1,000Yes$5.00 per 1,000NoYes$10.00 per 1,000Yes$12.00 per 1,000
WebhookThe results posted to your server when the search finishes.YesNoNoNoNo

Features on November 18, 2026

Read from each service's API reference. Parallel filters by a start date only. The deeper searches are querying.ai Pro Search, Parallel advanced, Tavily advanced and Exa Deep, at their lowest public prices.

API references

Parallel · Firecrawl · Tavily · Exa

One call returns pages your agent can read

Search finds the pages that answer a query and returns each one with its address, title, snippet, a relevance score and the passages that answer the query as plain text. Your agent gets its context from one request, with no second call to fetch the pages.

Search and Pro Search on one endpoint

type picks the search. A standard search answers in about three seconds. A pro search covers more of the web and picks passages more carefully for questions that need several sources, and answers in about four seconds.

Step 1

Send a query, read the results in the same response

Send query to POST /v1/search and the request stays open until results come back in the same response. numResults sets 1 to 10 results, and contents.text.maxCharacters sets how much page text each one carries, 200 to 10,000 characters.

cURL
curl -X POST https://api.querying.ai/v1/search \
  -H "Authorization: Bearer $QUERYING_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "query": "how do heat pumps work in cold climates",
    "type": "pro",
    "numResults": 3,
    "contents": { "text": { "maxCharacters": 500 } }
  }'
Step 2

Read the passages straight from the response

results arrive in score order. text holds the passages of each page that answer the query, so the agent passes them to its model as they are.

JSON
{
  "task": {
    "taskType": "SEARCH",
    "status": "COMPLETED",
    "latencyMs": 4063
  },
  "response": {
    "type": "pro",
    "query": "how do heat pumps work in cold climates",
    "results": [
      {
        "url": "https://www.carrier.com/us/en/residential/hvac-resources/heat-pumps/cold-climate-heat-pump/",
        "title": "Cold Climate Heat Pump Systems | Efficient Heating in Extreme …",
        "snippet": "In the winter, heat pumps work in cold weather by extracting thermal energy from the outdoor air and transferring it …",
        "text": "## Do cold climate heat pumps work?\n\nA heat pump is a versatile, year-round system that heats and cools your home by moving heat energy rather than generating it. In the winter, heat pumps work in cold weather by extracting thermal energy from the outdoor air and transferring it indoors.\n\n## Innovation in Action: The DOE Challenge\n\nCarrier is committed to pushing the boundaries of what a heat pump in cold weather can do. Carrier completed the Department of Energy’s Cold Climate Heat Pump Challen",
        "score": 1,
        "publishedDate": "2026-09-15T11:03:00.000Z"
      },
      {
        "url": "https://www.mitsubishicomfort.com/articles/mythbusters-heat-pumps-for-extreme-cold-climates",
        "title": "Heat Pumps in Extreme Cold Climates | Mitsubishi Electric HVAC US",
        "snippet": "In cold climates, heat pumps typically experience a decrease in pressure, refrigerant flow rate and operational …",
        "text": "## Myth: Heat Pumps Are Only for Mild and Warm Climates\n\nHow do they work? In cold climates, heat pumps typically experience a decrease in pressure, refrigerant flow rate and operational capacity. Variable-capacity heat pumps with Hyper-Heating INVERTER (H2i or H2i plus technology) allow for higher compressor speeds and pressures at lower outdoor air temperatures. This special function provides superior heat transfer performance.",
        "score": 0.8333,
        "publishedDate": "2026-10-19T12:06:06.108Z"
      },
      {
        "url": "https://homes.rewiringamerica.org/articles/heating-and-cooling/heat-pumps-cold-weather",
        "title": "Do heat pumps work in cold climates? - Rewiring America",
        "snippet": "This is why in extremely cold environments, your contractor will recommend a cold climate air-source heat pump — which generates …",
        "text": "In heating mode, heat pumps go into reverse and pump heat from the air outside your home to the inside. How does that work? Because heat is energy, there’s energy in the air even when it’s very cold outside. Believe it or not, super cold air — 0 degree Fahrenheit — contains more than 75 percent of the energy in 140 degrees Fahrenheit air!",
        "score": 0.6667
      }
    ]
  }
}

Shape each search with parameters

Only query is required.

  • type: standard or pro. Defaults to standard.
  • numResults: 1 to 10 results. Defaults to 5.
  • contents.text.maxCharacters: 200 to 10,000 characters of page text per result. Defaults to 2,000.
  • includeDomains, excludeDomains: up to 20 sites to search within or leave out.
  • userLocation: a two-letter country code for one market's results.

Fields in each result

results is sorted by score.

  • url, title, snippet: as the search engine's results page shows them.
  • text: the passages of the page that answer the query, in page order, with headings, lists and tables kept as plain text.
  • score: relevance from 0 to 1, comparable within one response.
  • publishedDate, siteName: present when the page declares them.

Read a whole page with Scrape

When a result needs more than its passages, send its address to POST /v1/async/task as taskType WEB_SCRAPE. It returns the page's main content as Markdown with its title, metadata and images, and reads PDFs too, for 0.3 credits a page: $0.75 per 1,000 pages on Basic, $0.30 on Max. MCP clients call it as the web_scrape tool.

0.5 credits a search, 2 a pro search

A search is charged when it completes, and a failed search is not charged. 1,000 searches cost $1.25 on Basic and $0.50 on Max; 1,000 pro searches cost $5.00 and $2.00. A new account starts with 500 free credits, which is 1,000 searches.

View pricing →
Request cost by plan
Plan1,000 requests
Basic≈ $1.25
Max≈ $0.50

Effective rates assume all included monthly credits are used.

Explore related capabilities

Frequently asked questions

How much do 1,000 searches cost?

A search costs 0.5 credits and a pro search 2. Per 1,000 searches that is Basic $1.25, Pro $0.59, Scale $0.52, Max $0.50; per 1,000 pro searches, Basic $5.00, Pro $2.35, Scale $2.08, Max $2.00.

When is a search charged?

When it completes. A failed search is not charged.

How fast is a search?

A standard search returns its results in about three seconds and a pro search in about four.

Is the text rewritten or summarized?

No. Every word in text comes from the page itself, so your agent can quote it and cite the address.

How is this different from the Google SERP API?

GOOGLE_SERP returns Google's results page as it is, with positions, ads and panels. Search returns the passages of the pages that answer the query, ready for an agent to read.

How do I read a whole page?

Send its address to Scrape (WEB_SCRAPE), which returns the page's main content as Markdown. A page costs 0.3 credits; per 1,000 pages that is Basic $0.75, Pro $0.35, Scale $0.31, Max $0.30.

Start with 500 free credits

Run your first task in the Playground, or call the API with your key.