The best web search API in 2026, by use case

Quick answer. There is no single best web search API; the right one depends on what your app does with the results. For AI agents that need the answer, not a list of links, browserfoundry by Keirolabs deep finds the answer to 96.5% of SimpleQA questions at $1.78 per 1,000 requests (Startup plan), the lowest price among APIs scoring above 93%. For the cheapest raw results, Keirolabs lite costs $0.03 to $0.10 per 1,000 on a plan (all 10 priced). For raw Google rankings, use Serper. For "find pages like this one", use Exa. Prices and scores below were checked in October 2026.

The best web search API for each job

The rest of this page shows how we got these numbers and where each pick falls short.

Accuracy vs price: the full comparison

SimpleQA is OpenAI's set of short factual questions with one verifiable answer each. Most search API vendors now publish a score on it, so it is the closest thing to a shared yardstick. A higher score means the answer is more often inside what the API returns.

How to read this table. Treat the top five as a tie: every vendor runs SimpleQA a little differently, and gaps under about a point are inside run-to-run noise. The price column is not noise. At this accuracy, Keirolabs deep is the only option under $5 per 1,000.

Sources: our score is from 200 SimpleQA questions none of which were used to build or tune deep (full write-up). Exa and Parallel are from You.com's harness, Perplexity from Tavily's harness, and the others are each vendor's own published number. Competitor prices are list pay-as-you-go rates, checked on October 7, 2026.

Best web search API for AI agents

An agent does not want ten blue links. It wants the passage that answers the question, a source to cite, and a signal for whether to keep looking. That is what deep returns:

• evidence: about 900 tokens of the best passages from the top pages, already de-duplicated across sites, each with its URL.
• answer_status: found, weak or not_found. When deep says found, the answer was in the evidence 97.3% of the time (183 of 188 in our test), so your agent can stop searching and move on.
• results: the ranked pages, each with answer_likelihood, an excerpt, and the full page text if you ask for it.

One call costs 2 credits on every plan. Search, page fetching, ranking and evidence packing are all included.

Best search API for minimizing token usage

Token cost usually dwarfs search cost. If your agent reads 10 full pages per question, that is often 30,000+ tokens of input before it writes a word. Here is what the answer rate looks like at different reading budgets, on the same SimpleQA questions:

The last two rows read the same ~900 tokens. Picking passages by keyword keeps the answer 75.6% of the time; deep's passage scoring keeps it 94.0%. That gap is the whole argument for an answer-layer API: you pay for one search call instead of 40x the tokens.

Best search API for citation accuracy

Citations fail in two ways: the cited page does not contain the claim, or the claim came from a page you did not cite. Two things reduce both:

1. Passage-level sources, not page-level. Deep attaches the URL to each evidence passage, and lists other sites that say the same thing under sources. Your model cites the passage it actually used.
2. A "not found" signal. An API that always returns something invites the model to cite a near miss. answer_status: not_found tells it there was no supporting passage, so it can say so instead of inventing a citation.

If you need citations from a fixed set of domains, run lite or fast with your own allow-list, then fetch only those pages with extract.

Best search API for low latency and real-time results

Measured end to end from a single client on October 30, 2026, with cold (uncached) queries:

Repeat queries are served from cache in under a second. If your hard requirement is a sub-second first byte for chat or voice, Parallel's Turbo mode (~200 ms, per its documentation) or Serper (~1 to 2 s) are the honest picks; use deep where accuracy matters more than a few seconds.

All Keirolabs search results are fetched live at query time, so breaking news shows up as soon as search engines index it.

Cheapest web search API

For raw results (title, URL, snippet), Keirolabs lite is the cheapest we know of: $0.10 per 1,000 on Essential, $0.06 on Pro, $0.03 on Startup. Serper is the cheapest Google-specific option at $1.00 per 1,000, falling to $0.30 only after you prepay for 12.5 million queries.

Cheap per request is not the same as cheap per answer. Lite alone has the answer in its snippets 59.4% of the time; to get to 90%+ you add crawling and model tokens. If your agent needs answers, compare deep at $1.78 against lite plus your own crawler plus the tokens to read it all.

Best free web search API

Bing's Search API was retired on August 11, 2025, and Google's Custom Search JSON API is closed to new signups, which is why most free search now comes from the providers above.

Every Keirolabs endpoint and what it costs

One API key covers every endpoint. Credits cost $2.67 per 1,000 on Essential ($29.99/month) and $0.89 per 1,000 on Startup ($99.99/month).

Full plan details are on the pricing page.

Where Keirolabs is not the best choice

• You need Google's exact ranking. Use Serper or SerpAPI. Our results come from several engines and are re-ranked, so they will not match Google position for position.
• You need sub-second responses. Lite and fast take seconds, not milliseconds. Parallel Turbo or Serper are faster.
• You need semantic "find similar" retrieval. Exa is built for that; we are keyword-and-intent first.
• You need a compliance package on day one. If a SOC 2 report is a hard procurement gate, check which vendors can hand one over today before you shortlist.
• Some sites block all automated readers. Deep marks these pages blocked instead of guessing. Medium is a common one.

How to call it

curl

curl -X POST https://api.keirolabs.cloud/api/v2/deep \
  -H "Authorization: Bearer keiro_..." \
  -H "Content-Type: application/json" \
  -d '{"query": "how many moons does neptune have"}'

Python

import requests

r = requests.post(
    "https://api.keirolabs.cloud/api/v2/deep",
    headers={"Authorization": "Bearer keiro_..."},
    json={"query": "how many moons does neptune have"},
    timeout=90,
)
data = r.json()
print(data["answer_status"])             # found / weak / not_found
for e in data["evidence"]:
    print(e["url"], e["text"][:200])

TypeScript

const res = await fetch("https://api.keirolabs.cloud/api/v2/search/lite", {
  method: "POST",
  headers: { Authorization: "Bearer keiro_...", "Content-Type": "application/json" },
  body: JSON.stringify({ query: "best vector database for rag" }),
});
const { results } = await res.json(); // [{ title, url, snippet }, ...]

MCP (Claude, Cursor, any MCP client)

{
  "mcpServers": {
    "keirolabs": { "type": "http", "url": "https://api.keirolabs.cloud/mcp" }
  }
}

Get a key at platform.keirolabs.cloud. Keys start with keiro_, and the free credits work on every endpoint.

How we tested

• Accuracy: 200 questions from OpenAI's SimpleQA set, none used while building deep. A question counts as answered when the reference answer appears in the returned text. The lite numbers use 500 questions and an open harness you can rerun.
• Prices: each provider's public pay-as-you-go page, checked October 7, 2026. Volume discounts and enterprise deals are not included.
• Latency: wall-clock time for cold queries from one client on October 30, 2026. Your numbers will vary with region and query.

We sell one of the APIs on this page. Every competitor number above is either the vendor's own or from a third party's harness, and the sources are named so you can check them.