Best Free Web Search APIs for LLM Agents: 3 Tested Live

Does Google Have a Free Search API?

Not for new projects. Google's Custom Search JSON API overview says the API "is closed to new customers", and existing customers have until January 1, 2027 to move to an alternative. For full web search, the page asks you to contact Google. The usual route to Google results for an agent is a third-party API, and the three below each have a free tier that we ran live.

Quick Answer: Best Free Web Search APIs at a Glance

Serper, OpenWeb Ninja and SerpApi all return Google results as JSON, with a free tier and no credit card. All 45 of our October 2026 calls returned results. In our run, checked against official sources, the top 3 results most often held the current answer with Serper (14 of 15 tries), then OpenWeb Ninja (12) and SerpApi (8). OpenWeb Ninja had the fastest median response, 0.57 s.

Records are organic results per call. "Correct" counts answers from the top 3 results that matched the official source, over 5 questions and 3 runs. The order follows correct answers, then median response time. We build OpenWeb Ninja (#2), so read its row with that in mind.

How We Evaluated

• Scope: search APIs that return Google results, with a free tier, tested live with our own keys. Search APIs that do not return Google results, including those built specifically for LLMs, are out of scope: this test compares the same Google results across providers.
• Five questions an agent gets asked: the latest stable version of Python, who won the 2026 World Cup, the latest Node.js LTS version, the current federal funds target range, and who the current CEO of OpenAI is.
• Live calls, October 2026: each question three times on each API, 45 calls in all, in one session from one machine.
• Matched settings: US results in English (gl=us, hl=en) on all three. OpenWeb Ninja and Serper were asked for 10 results. SerpApi's docs list no result-count parameter, so it returned Google's default page, with no_cache=true so every run was a fresh search.
• Measured: round-trip response time, organic results per call and response size, from the saved responses.
• Grounding check: for each question and run, Claude Opus 5.5 (Anthropic) read only the title, URL and snippet of the top 3 results and gave an answer with its source. We scored each answer against the official source (python.org, fifa.com, nodejs.org, federalreserve.gov, openai.com) as correct, out of date or not found, comparing the same figure in every column.
• Order: correct answers first, then median response time.
• Sample size: five questions, three runs each: a smoke test, not a benchmark. A different question set could change the order.
• Pricing: read from each provider's own pricing page in October 2026.

We build OpenWeb Ninja, so read its section with that in mind. Everything below for the other two comes from our own calls and their own pages.

Best for: fast, compact Google results for an agent, on monthly plans, with 50+ other data APIs on the same key

The Real-Time Web Search API returns Google results in real time. Its light endpoint, search-light, returns organic results only, which keeps the response small enough to drop straight into a prompt. The code below runs as shown with your key and prints the top 3 results as model context.

const params = new URLSearchParams({ q: "who won the 2026 world cup", num: "10", gl: "us", hl: "en" });
const res = await fetch(`https://api.openwebninja.com/realtime-web-search/v2/search-light?${params}`, {
  headers: { "x-api-key": "YOUR_API_KEY" },
});
const { data } = await res.json();

// Top 3 results, ready to paste into the model's context
const context = data.slice(0, 3).map(({ title, url, snippet }) => ({ title, url, snippet }));
console.log(JSON.stringify(context, null, 2));

Response from the benchmark (World Cup question, first run), trimmed

{
  "status": "OK",
  "request_id": "938d8a97-f2ea-4738-a2e7-bfb92398064d",
  "parameters": { "q": "who won the 2026 world cup", "num": 10, "start": 0, "gl": "us", "hl": "en" },
  "data": [
    {
      "title": "FIFA World Cup 2026™ Final",
      "snippet": "Spain claimed ultimate glory in front of a crowd of over 80,000 at New York New Jersey Stadium, as they were crowned FIFA World Cup™ champions on Sunday 19 July ...",
      "url": "https://www.fifa.com/en/tournaments/mens/worldcup/canadamexicousa2026/final",
      "domain": "www.fifa.com",
      "position": 1
    },
    {
      "title": "FIFA World Cup",
      "snippet": "The reigning champion is Spain, who won their second title at the 23rd edition of the competition, the 2026 World Cup, by defeating the holders, Argentina.",
      "url": "https://en.wikipedia.org/wiki/FIFA_World_Cup",
      "domain": "en.wikipedia.org",
      "position": 2
    },
    ... 6 more results
  ]
}

• Free tier: 100 requests a month, hard limit, 1 request per second, no credit card.
• Paid plans: Pro $25 for 10,000 requests, then $0.003 each. Ultra $75 for 50,000. Mega $150 for 200,000. Pay As You Go $0.005 per request with no monthly fee.
• Our run: 15 of 15 calls returned results: 7 to 10 per call, median 0.57 s (0.47 to 0.82 s), the fastest here, and a median response of 3.1 KB.
• Grounding check: the Fed question was the gap: the top 3 held the effective rate and a 4 percent figure, but not the target range.
• Full search: the search endpoint adds the total result count and an AI Overview flag. It costs 2 requests per call, which we confirmed on our usage meter.
• For agents: an MCP server exposes the API as a tool, and the API has an llms.txt for agents that read docs. For answers from Google AI Mode, see the Google AI Mode API.

Can an LLM Answer From the Results? Opus 5.5 Grounding Check

One thing worth checking when you pick a search API for grounding AI answers is whether the top results already hold the current answer and a source to cite. This check measures that.

Claude Opus 5.5 read only the top 3 results (title, URL, snippet) from each of three runs in October 2026. Each answer was scored against the official source shown.

• All three APIs led to Spain in all three World Cup runs. In SerpApi's runs, the #1 result was a fifa.com page about past winners, and one run included an Al Jazeera snippet from before the tournament calling Argentina the reigning champion. Britannica, also in the top 3, named Spain each time.
• Out-of-date answers came from older pages ranking in the top 3: Python 3.12 from a Reddit thread (Serper, one run), Python 3.13.0 and Node.js 24.11.0 (SerpApi), and 2023 news about OpenAI's CEO (SerpApi, one run).
• On the Fed question, OpenWeb Ninja's top 3 held the effective rate (3.88 percent, consistent with the current range) but not the target range itself, so we scored it not found.
• Why the top 3 differ: the three APIs did not return the same top 3 for the same query and settings, and the same API's top 3 also changed between runs on 10 of 15 question and API pairs. OpenWeb Ninja and Serper returned an identical top 3 in 5 of 15 runs. SerpApi returns Google's default page with no result-count parameter, and where its answers were out of date, python.org and nodejs.org pages sat lower in its list (python.org at positions 6 and 8, nodejs.org at 5 to 9). We did not measure the cause beyond that.
• SerpApi's responses hold more than the top 3: related questions in all 15, and AI Overview text inline in 2 (10 more pointed to a separate request). An agent that reads those has more to work with than this check measured.

How to Choose: 3 Common Scenarios

To wire any of these into a model, see Give Your AI Agent Live Web Search. For SEO-focused SERP APIs on paid plans, see Best SERP APIs in 2026, and the head-to-head pages OpenWeb Ninja vs Serper and OpenWeb Ninja vs SerpApi. More use cases: AI, RAG and LLM grounding.