Exa vs Tavily vs Firecrawl vs Parallel: the web search APIs for AI agents, compared

Exa vs Tavily comes down to the job: Exa searches its own index of 100 billion documents by meaning, and Tavily pairs fast search with page extraction in one credit-priced API. Firecrawl turns any page or site into clean markdown, and Parallel adds deep research and monitors on its own index, all four reading the public web. Obsession gives agents what a company shows only its signed-up customers, such as the welcome email and member price, each dated with its screenshot, by API or MCP (the standard way agents call tools).

Exa, Tavily, Firecrawl and Parallel each give an AI agent the web in one API call, and each is strongest at a different job. Exa searches its own index by meaning, and Tavily pairs fast search with page extraction. Firecrawl turns any page or whole site into clean markdown, and Parallel adds deep research and scheduled monitors on its own index.

For developers and AI engineers choosing one, six differences decide it. They are where results come from, how fresh they are, how much page text comes back, how fast a call returns, what 1,000 searches cost, and how an agent connects over MCP, the standard way AI agents call tools.

All four share one boundary. They read the public web, the pages any visitor can open, plus the licensed and partner data some of them add. What a company sends and shows a customer after the sign-up, such as the welcome email, the code inside it and the signed-in price, sits past that boundary.

How the four were compared

Every fact about a provider comes from its own pricing, product and documentation pages as of 11 November 2026. Seven criteria separate them.

• Index: whether the API searches a copy of the web it crawled and stores itself, or passes the query to another search engine. An index is that stored, searchable copy, kept current by a crawler that fetches pages on a schedule.
• Freshness: how old a page can be when it reaches the agent, and the setting that forces a new fetch.
• Page content: links and snippets, the passages that answer the query, or the whole page as markdown or JSON.
• Speed: the time a call takes, as each provider states it. A p50 is the median call; a p95 means 95 calls in 100 finish faster.
• Price per 1,000: the list price of 1,000 searches of 10 results, worked out from credits where a provider prices in credits.
• Free tier and MCP: what each gives free every month, and whether an agent such as Claude or Cursor can connect with no code.
• Behind the login: whether its pages describe reaching what a site shows only after a sign-in, and keeping it over time.

Exa: its own index, searched by meaning

Exa calls itself “The world’s data for AI agents” and says it searches “the largest index of public web and private information”. Its search page describes Exa’s own index of 100 billion documents, with 3.5 billion pages crawled every day. Exa searches it by meaning as well as by keywords, so it finds pages that never use the query’s exact words.

The same page pairs the public web with “proprietary data”: dedicated data on 70M+ companies, 1B+ people and 211M+ places, plus research papers, code and financial filings. Partner data comes through Exa Connect, from providers such as Crunchbase, ZoomInfo and Similarweb.

• Search types: Lite in about 450 ms, Auto in about 1 second, and Deep in 4 to 15 seconds for multi-step research with structured outputs and citations.
• Page content: text and highlights come back in the search response at no extra cost. Exa says its highlights and structured outputs cut an agent’s input tokens by 95% on average.
• Freshness: set contents.maxAgeHours and Exa fetches a new copy of any page whose stored copy is older than that limit.
• Beyond search: Contents returns full pages, Answer returns an answer, and Agent handles long research runs, list building and enrichment. Monitors run a search on a schedule and post updates to a webhook, a message sent to your server when something new appears.
• Price: its pricing page lists Lite at $4, Auto at $7 and Deep at $12 per 1,000 requests with up to 10 results. Contents costs $1 per 1,000 pages and an Agent run $0.012 to $1.00. The free plan gives $10 of credit every month plus a $10 onboarding bonus, with no payment method.
• MCP: a hosted server at mcp.exa.ai, free and with no API key, as Exa’s MCP page says.
• Where it stops: its pages describe searching the public web, plus dedicated and partner data. They don’t describe signing up to a company or keeping what a company sends its customers.

Tavily: search and extraction built for agents

Tavily describes itself as “One secure API for real-time web access”, with search, extract, crawl and research endpoints, and gives its median search time as 180 ms. A search returns ranked results with snippets, and can add each page’s cleaned content as markdown and an answer written by a language model in the same call.

• Search depth: ultra-fast, fast and basic each cost 1 credit, and advanced costs 2 for the highest relevance at more latency, as its search reference describes. A search returns at most 20 results.
• Freshness: results filter by publish date, as a time range or start and end dates, and a news topic. The same reference warns that “Continuously updated pages, such as weather forecasts and prices, keep old publish dates and get filtered out”, so a date filter can drop the pricing page an agent needs.
• Extraction and crawling: Extract costs 1 credit per 5 pages, or 2 at advanced depth, and Crawl adds the cost of mapping a site’s links, per its credits page. A Research request costs 4 to 110 credits on mini and 15 to 250 on pro.
• Price: its pricing page gives 1,000 free credits a month with no payment card, and $0.008 a credit pay as you go. Monthly plans run from $0.0075 a credit down to $0.005 on the $500 Growth plan.
• MCP: a remote server with search, extract, crawl, map and research tools, and keyless access to search and extract, per its MCP docs.
• Ownership: Nebius announced an agreement to acquire Tavily on 10 February 2026, and Tavily’s own post says the API, its data policies and zero data retention stay the same.
• Where it stops: its pages describe the public web. They don’t describe signing in to a site or keeping what a company sends its customers.

Firecrawl: any page or whole site as clean markdown

Firecrawl calls itself “The infrastructure layer that helps AI find, read, and act on the live web.” Its core job is extraction. Give it a URL and it returns the page as markdown, HTML, JSON in a schema you define, or a screenshot, and its crawl endpoint follows links across a whole site. The code is open source, and the hosted version adds Firecrawl’s own proxy and rendering engine.

• Search: results can come back with each page’s full content. Its search docs refer to “our upstream search provider”, so its search draws on another provider’s results before Firecrawl scrapes them.
• Freshness: a scrape is served from cache when a copy under 2 days old exists. Setting maxAge to 0 forces a live fetch, which its scrape docs say takes longer and fails more often.
• Speed: a p95 of 3,387 ms on its own benchmark of 1,000 URLs, run on 13 January 2026.
• Interact and logins: Interact keeps a browser open across calls, driven by plain-language prompts or code. A login lives in a session cookie, a token the site sets at sign-in and checks on every request. A saved profile of cookies and local storage lets a later session start signed in. Its Interact docs describe signing in once and then reading the pages that appear only after authentication. Sessions end after 10 minutes by default, or 5 minutes idle, and cost 2 credits a browser minute, or 7 when driven by a prompt.
• Price: its pricing page gives 1,000 free credits a month with no payment card. A scrape or crawl costs 1 credit a page and a search 2 credits per 10 results. Standard is $99 a month, or $83 billed yearly, for 100,000 credits, which Firecrawl counts as 50,000 searches or 100,000 pages.
• MCP: an official server for Cursor, Claude, Windsurf and other clients, to search, scrape and interact, per its home page.
• Where it stops: a session lasts minutes, on an account your team signs in to, and its pages don’t describe receiving the emails or texts a sign-up sets off. For your own accounts, how to scrape data behind a login covers sessions, cookies and one-time codes.

Parallel: its own index, deep research and monitors

Parallel builds “Web infrastructure for AI to search, extract, monitor, and reason over the world’s information” and runs “our own web-scale index with billions of pages and millions more added and updated daily”. The index includes licensed content from paywalled publishers, premium newsletters and other sources through its Index partnerships, launched on 19 May 2026, and a live crawl can be forced for time-sensitive queries.

• Search modes: turbo in about 200 ms, fast in about 700 ms, basic in about 1 second, and advanced, the default, in about 3 seconds, per its modes docs.
• Beyond search: Extract reads PDFs, pages built in JavaScript and government sites. Task runs deep research and enrichment in 10 seconds to 2 hours, and FindAll builds verified lists. Monitor watches the web hourly, daily or weekly and posts each change to a webhook, per its Monitor docs.
• Price: its pricing docs list Search at $1 per 1,000 requests in turbo or fast mode and $5 in basic or advanced, with 10 results included. Extract is $1 per 1,000 URLs, Task $5 to $2,400 per 1,000 runs and Monitor $3 to $10 per 1,000 checks. Its pricing page adds $5 free a month with a payment card on file, plus up to $80 at signup.
• MCP: a hosted server at mcp.parallel.ai, whose search and fetch tools are free at lower rate limits, per its MCP page.
• Where it stops: its licensed sources are publishers who license their work. Its pages don’t describe signing up to a company to receive what it sends its customers.

Index, freshness and speed

The freshness settings differ more than the headlines do. Exa refetches any page older than the maxAgeHours you set, Firecrawl serves a cached copy up to 2 days old unless maxAge is 0, and Parallel can force a live crawl. Tavily filters by publish date, which removes pages that change in place, such as prices. For a question about this month’s prices, fetch the page itself or force a refresh, and leave publish-date filters for news.

Price per 1,000 searches

Small gaps per 1,000 grow fast at agent volumes. Between $1 and $4 per 1,000 searches, the gap is $300,000 a month at 100 million searches, $30,000 at 10 million and $3,000 at 1 million. Those volumes are examples; put your agent’s own searches per month in their place.

Price per search is half the sum. A provider whose results answer the question in 1 call costs less than a cheaper one that needs 3. A response that carries the page text also saves a second call to an extraction endpoint. Exa’s highlights, Tavily’s raw content, Firecrawl’s scrape options on search and Parallel’s excerpts all exist to save that second call.

Benchmarks, and how to run your own

Each provider publishes a benchmark it leads. Exa’s search page says it ranks first for quality on the Artificial Analysis Search Index and charts Exa Lite ahead of Parallel, Brave and Perplexity. Tavily’s home page puts it first on SimpleQA and SealQA, and Parallel’s charts accuracy against cost on OpenAI’s BrowseComp. Your agent’s own questions are the fair test, and the four free tiers cover one.

1. Take 50 real queries from your agent’s logs, the hard ones included.
2. Write the answer you expect for each, with the page that proves it.
3. Run every query through each free tier, with the same number of results and the same model reading them.
4. Score each answer right, partly right or wrong, and record the latency and cost of each call.
5. Run the queries about prices and news again a week later, to see which provider kept up.

Exa’s first month of free credit, with its onboarding bonus, covers about 5,000 Lite searches, Tavily’s 1,000 basic searches, Firecrawl’s 500 searches and Parallel’s up to 5,000 requests, so 50 queries fit every one.

One question, worked through

Say an agent is asked what Quillon, an invented online shop at quillon.example, offers new subscribers this month. Each API answers from what it can reach.

• Exa, Tavily or Parallel: Quillon’s home page, which promises 10% off for signing up, and perhaps a coupon site repeating it. The agent reports the promise.
• Firecrawl: the same promise as clean markdown or a screenshot, or a click through the sign-up form inside an Interact session that ends within minutes.
• What a new subscriber got: a welcome email with a 15% code 4 minutes after signing up, and a second email 3 days later offering free delivery over £30. After signing in, the price was £24, against £28 on the public page.

The welcome email went to an inbox, and the second arrived days after any browser session had ended. A search index stores pages and holds no inbox, and a scraper reads what a page shows at the moment it visits. Firecrawl could read the £24 price by signing in to a Quillon account your team holds, but the emails never reach a page. So the honest answer from the four APIs is the 10% the page promises, and the 15% code goes unseen.

What a company shows only its customers

Obsession collects what companies show only to customers, step by step: their screens, emails and texts, member prices, trial limits, quotes and cancel offers, each dated with a screenshot. Teams and AI agents look up any company through the API or MCP, and a company not yet covered is added on request.

Exa, Tavily, Firecrawl and Parallel give agents the public web, what any visitor can read. Obsession gives them what a company shows only to its customers, behind the login, dated and kept for as long as anyone follows the company. Search for what a company publishes and look up Obsession for what it sends and shows its customers; an agent can call both in the same answer.

In the Quillon example, a call to facts for quillon.example returns the welcome email with its 15% code, the delivery offer and the £24 signed-in price. Each comes with the time it arrived and the email or screenshot behind it. Obsession signs up the way any customer can and keeps each account open for months. Obsession’s agents have run 1,500 sign-ups, baskets and checkouts at real companies, and counting. The sample data shows one real store’s record, and the web behind the login explains what sits behind each step.

The API and MCP expose it as 5 verbs: lookup, facts, evidence, index and notify. TypeScript and Python SDKs call the same verbs, and a webhook delivers each new fact. It is priced per call or per company watched, never per token, and every key has a spending cap you set.

Exa, Tavily, Firecrawl and Parallel compared

We make Obsession, so every other option is described from its provider’s own pages as of 11 November 2026.

Which API fits which job

Use Obsession when your agent needs what a company sends and shows only its customers, at any company you name. That covers the welcome series, member and signed-in prices, trial limits and cancel offers, each dated with the email or screenshot behind it and kept for as long as you follow the company. For the public web, pick by job.

• What any company sends and shows its customers, dated, by API or MCP: Obsession.
• Search by meaning across the public web, plus company and people data: Exa.
• The fastest grounding call in an agent loop: Parallel’s turbo mode at about 200 ms and $1 per 1,000, or Tavily at a 180 ms median.
• Search and page text in one credit-priced API: Tavily.
• Clean markdown or JSON from any URL, a whole-site crawl, or a page behind a login your team owns: Firecrawl.
• Deep research reports, verified lists and scheduled monitors: Parallel’s Task, FindAll and Monitor, or Exa’s Agent and Monitors.
• A one-off question about 1 company: a browser and a spare inbox, by hand.

Exa alternatives sorts more providers by job, from low-cost search result APIs to extraction tools. The best web search APIs for AI agents prices the wider field of grounding sources beside these four, and the comparison page sets out every other source, from email archives to panels.

Before you commit

• Zero data retention, where the provider keeps no copy of your queries or results, matters when queries carry customer details. Exa offers it on Enterprise with a $12,000 yearly commitment, and Parallel includes hybrid zero data retention on its $1,000 a month Scale plan, by signed agreement. Firecrawl offers it on Enterprise for 1 extra credit a page, and Tavily calls it core.
• Rate limits: Parallel allows 600 search requests a minute on Build, and Firecrawl 10 scrapes a minute free and 500 on Standard.
• Plans change often. Check each pricing page on the day you buy.

Frequently asked questions