TL;DR
- Lowest cost: Keirolabs lite, $0.10 per 1,000 ($0.03 on Startup).
- Highest accuracy you can verify: Keirolabs deep (96.5%) and You.com (96.11%), both with code and results in public repos.
- Page text for RAG: Exa. Fast and cheap for agents: Parallel at $1 per 1,000.
- Independent index: Brave. Google's exact results: Serper.
- Most accuracy numbers are self-reported. The same API can score 76% or 94% depending on who runs the test.
Ask ten developers for the best web search API and you’ll get ten answers, nine of which are the API they happened to try first.
That’s not because they’re wrong. It’s because “best” depends entirely on the job. The best API for a rank tracker is a bad fit for a RAG pipeline. The best one for a chatbot your users read is a poor fit for an agent loop that fires 400 searches per task. The cheapest one might be the most expensive once you count the tokens it makes your model read.
So this guide doesn’t crown a winner. It gives you a winner per job, with real prices (checked November 2, 2026) and an honest look at the accuracy claims, which are messier than any vendor’s landing page suggests.
The short list
| If you need | Pick | Per 1,000 | Why |
|---|---|---|---|
| Lowest cost | Keirolabs lite | $0.10 ($0.03 on Startup) | Own index, lowest list price we track |
| Highest verifiable accuracy | Keirolabs deep · You.com | $1.78–5.33 deep · $5 | 96.5% and 96.11% SimpleQA, code and results public |
| Page text for RAG | Exa | $4 lite · $7 auto | Own neural index, contents $1 per 1,000 pages |
| Fast and cheap for agents | Parallel | $1 fast · $5 basic | Ranked, compressed excerpts save tokens |
| Independent index | Brave | $5 | One of few Western independent indexes at scale |
| Google’s exact results | Serper | $0.30–1.00 | Cheapest Google SERP JSON |
| One line in your model call | OpenAI / Claude web search | $10 + tokens | Built in, but locked to that model |
Not sure which job you have? The picker asks three questions and gives you a shortlist. Know your volume? The calculator prices every mode.
First, the uncomfortable part: accuracy claims
Every search API vendor quotes a SimpleQA score. SimpleQA (OpenAI, 2024) is a set of short factual questions with one correct answer each, and it has become the benchmark everyone points at.
The catch: almost every vendor quotes their own run, using their own harness, their own model to read the results, and their own grading. Then they test their competitors the same way. You can guess how that usually turns out.
| API | SimpleQA | Measured by | Can you check it? |
|---|---|---|---|
| Keirolabs deep | 96.5% | Keirolabs | ✅ Public repo |
| Keirolabs lite + crawl recipe | 92.0% | Keirolabs | ✅ Public repo |
| You.com | 96.11% | You.com | ✅ Public repo |
| Exa (search + text) | 95.40% | You.com | ✅ In You.com’s repo |
| Parallel (basic) | 94.87% | You.com | ✅ In You.com’s repo |
| Tavily | 93.3% | Tavily | ✅ Public repo |
| Tavily | 88.53% | You.com | ✅ In You.com’s repo |
| Perplexity Search | 85.92% | Tavily | ✅ In Tavily’s repo |
| Brave | 76.05% | Tavily | ✅ In Tavily’s repo |
The same API can score 76% or 94% depending on who’s holding the stopwatch.
Brave is the clearest example: 94.1 F1 in Brave’s own AI Grounding test, 87% in Parallel’s, 76.05% in Tavily’s. That’s an 18-point swing for one product. Tavily scores itself 93.3% and gets 88.53% in You.com’s harness.
Treat SimpleQA numbers as “this API is in the conversation”, not as a ranking. Above roughly 90%, the differences between top APIs are smaller than the differences between test setups. Your own 100 queries will tell you more than any of these tables, including ours. The full table has 23 claims, each linked to its source.
The picks, one by one
Keirolabs: the lowest cost
$0.10 per 1,000 lite searches on Essential, $0.03 on Startup. That’s 10× under Serper’s entry price and 50× under most AI search APIs. It runs its own index, and the same key unlocks fast search, page content, answers, a deep research mode and a Google SERP endpoint, all priced in credits. It ships an MCP server for Claude Code, Cursor and friends.
- Free tier: 750 credits, once
- SimpleQA: 96.5% on deep (public repo, 200 questions) and 92.0% for lite plus an open crawl recipe (public repo, 500 questions). Lite’s titles and snippets alone hold the answer 59.4% of the time. Both runs check whether the answer is in what comes back, not whether a model gets it right.
- Pick it if you make a lot of calls and price per call decides it.
You.com: the most verifiable accuracy
You.com did something unusual: it published its benchmark harness with results for itself and its competitors, in one public repo. Its search scored 96.11%, and it’s the only public repo that runs competitors through the same harness. Search and the Answer API both cost $5 per 1,000. Research runs from $12 (lite) to $1,200 (frontier) per 1,000.
- Free tier: $10 of credit (about 2,000 searches)
- Pick it if accuracy matters most and you want numbers you can audit.
Exa: page text for RAG
Exa runs its own neural index, built on embeddings rather than keywords. It’s very good at “find pages like this” and fuzzy topic queries. Lite $4, auto $7, deep $12 per 1,000, and page contents for $1 per 1,000 pages. It scored 95.40% in You.com’s public run.
- Free tier: $10 of credits a month
- Pick it if your model needs clean page text and you’d rather not run a scraper.
Parallel: fast and cheap for agents
Parallel returns ranked, compressed excerpts instead of whole pages, so your model reads fewer tokens per answer. In an agent loop, that saving can matter more than the search price. Turbo and fast are $1 per 1,000, basic and advanced $5. Basic scored 94.87% in You.com’s run.
- Free tier: 5,000 requests a month
- Pick it if you run agents in loops and token bills are real.
Brave: the independent index
Brave crawls the web itself and never falls back to Google or Bing. For a product people read, that independence is the point: no one upstream can pull the plug. $5 per 1,000, including its LLM context endpoint. How to get a key.
- Free tier: $5 of credits a month (about 1,000 searches), card required
- Pick it if people read the results, or Google/Bing dependency is a deal-breaker.
Serper: Google’s exact results
Sometimes you need exactly what Google shows. Serper is the cheapest real-time way to get it: $1.00 per 1,000 on the small pack, $0.30 at 12.5M. It’s scraped Google, so the vendor carries the legal risk today. More in SERP API pricing and best SERP APIs.
- Free tier: 2,500 queries at signup
- Pick it if you need Google’s ranking specifically.
OpenAI and Claude web search: the one-liner
Both model vendors ship search as a built-in tool for $10 per 1,000 calls plus tokens. It’s the fastest thing to turn on. But the results count as input tokens on every later turn, you can’t swap the model and keep the search, and you don’t control ranking.
- Pick it if you’re prototyping, or your volume is small and simplicity wins.
The hidden cost: tokens
Here’s the number most comparisons skip. If an LLM reads your search results, you pay twice: once for the search, then again for the tokens the results add to the prompt, often on every later turn of the conversation.
Say each search returns about 3,000 tokens of page text, and your model charges $3 per million input tokens. That’s about $9 per 1,000 searches just to read the results. Then if the results stay in context for five more turns, multiply. Next to that, the difference between a $1 and a $5 search API is small. This is why compressed excerpts (Parallel) and tight snippets can beat a cheaper API that dumps whole pages.
Is there a free web search API?
Not one to run production on. The best free allowances right now:
| API | Free allowance |
|---|---|
| Parallel | 5,000 requests a month |
| Exa | $10 of credits a month |
| Brave | $5 of credits a month (≈1,000 searches) |
| Tavily | 1,000 credits a month, no card |
| Serper | 2,500 queries at signup |
| Keirolabs | 750 credits at signup |
DuckDuckGo has no real search API, and Bing’s was retired.
How to choose in one afternoon
Every comparison you’ll read, including this one, was written without your queries. Fix that:
- Pull 50–100 real queries from your logs.
- Run them through two or three of the picks above, on the same day.
- Grade answerability: is the answer in the top 5 results?
- Measure p50 and p95 latency with calls running in parallel.
- Count tokens the results add to your prompts.
- Divide cost by useful results. That’s the number that matters.
The full checklist is on how to choose. Then come back and tell us what you found.