Most search engines are built for people. You type a question, you skim ten blue links, you click the one that looks right. Exa 3.0 flips that idea on its head: it is a search engine built for software. Released on October 18, 2026 by the San Francisco startup Exa Labs, Exa 3.0 is a major upgrade to the company’s web search API for AI agents — and it arrives with a price cut, a bigger index, and a bold claim that has the whole developer world talking.
Here is the headline in one sentence: Exa 3.0 claims to beat every rival — Perplexity, Brave Search and Parallel — on all eight benchmarks it ran, while a brand-new Exa Lite tier brings the cost down to just $4 per 1,000 requests. For teams running AI agents that search the web hundreds or thousands of times a day, that price drop matters as much as the benchmark scores. Search has quietly become an infrastructure bill, and Exa wants to own that line item.
In this guide, we break down everything that changed in Exa 3.0: the benchmark claims and why you should read them with one eyebrow raised, the new Dynamic Highlights feature that cuts token use by 95%, the full Lite vs Auto pricing breakdown in plain USD, how the expanded index works, and how Exa stacks up against Perplexity, Brave and Tavily. We will also give you an honest verdict on whether Exa 3.0 deserves a place in your AI stack — and who should probably skip it.

Table of Contents
- What Exactly Is Exa 3.0?
- The Benchmark Claim: #1 on 8 Tests — With an Asterisk
- The Engine Room: Exa’s Largest Embedding Model Yet
- Dynamic Highlights: 95% Fewer Tokens, Better Answers
- Exa Lite vs Exa Auto: The Pricing Breakdown
- A Bigger, Fresher Index: What’s New in the Data
- Exa 3.0 vs the Competition
- Who’s Actually Using Exa?
- Why AI Agents Need Their Own Search Engine
- Getting Started With Exa 3.0
- The Honest Verdict
- Exa 3.0 FAQ
What Exactly Is Exa 3.0?
Exa is a startup with one big idea: as AI agents do more and more of our searching for us, search itself needs to be rebuilt for machines, not humans. A person wants a synthesized answer in a chat window. An agent — a coding assistant, a research bot, a shopping copilot — wants raw, structured, citable results it can feed into its own reasoning. That is the gap Exa 3.0 is designed to fill.
The company runs its own web crawler and its own independent index. This matters more than it sounds. Many “AI search” products are really wrappers around Google or Bing results, which means they inherit those engines’ ranking logic — logic tuned for human eyeballs and ad clicks. Exa 3.0’s results come from infrastructure the company built itself: it tracks more than 2 trillion URLs and serves 100 billion pages, and it grows the index around what its customers actually search for. If your agent keeps asking about, say, local business hours or product listings, the index leans into that coverage.
Exa 3.0 is not a website you visit. It is an API — developers call it from their code, and it returns search results, page contents and structured data that an AI model can read directly. Exa also ships related endpoints: a SERP API, a crawling API, an extracting API and a deep-research API. The 3.0 release upgrades retrieval quality across every one of these search types.
The company behind it, Exa Labs, was founded by CEO Will Bryk and is based in San Francisco with 51–200 employees. It was recently named to the Forbes Agentic 20, the inaugural Forbes list of companies leading agentic AI. That is a strong signal of where the industry thinks Exa belongs: not as a consumer search brand, but as plumbing for the agent economy.
The Benchmark Claim: #1 on 8 Tests — With an Asterisk
Exa 3.0’s launch announcement leads with a punchy claim: first place on every benchmark it tested, from code search to financial research, against rivals including Perplexity, Brave Search and Parallel. The numbers Exa published are specific — and specificity deserves a closer look, because it tells you exactly what “better search” means here.
Here are the scores Exa 3.0 reported for its top-tier Auto search type:
| Benchmark | What it tests | Exa 3.0 score |
|---|---|---|
| BrowseComp | Tracking down hard-to-find facts across the open web | 91.8% |
| DSQA | Factual question answering | 88.2% |
| WideSearch | Broad web search coverage | 56.1% |
| ATLAS | Agentic task evaluation | 42.0% |
| WebCode | Code search | 75.5% |
| FinSearchComp | Financial research search | 68.0% |
The company also reported leading scores on people-lookup and company-lookup tasks, which rounds out the eight.
Now for the asterisk — and it is an important one. These benchmarks were run by Exa itself. As outside coverage noted, Exa did not publish raw outputs or an independently reproducible test framework, so you should treat the rankings as a company claim rather than a referee’s verdict. That does not make the numbers fake; it makes them marketing-grade evidence, and a smart buyer weights them accordingly. Independent third-party benchmarks will be the real test, and they have not landed yet.
Still, the scoreboard is more interesting for its shape than its headline. BrowseComp at 91.8% is the standout: it tests an agent’s ability to hunt down obscure facts across the open web, which is precisely the job Exa 3.0 is sold for. ATLAS at 42.0% is the weakest score — and it still beat Perplexity’s 37.2%. These are not “we are ten times better” gaps; on BrowseComp the lead over Perplexity was reportedly just 0.3 points. The honest reading: Exa 3.0 is incrementally ahead on the tests it chose to publish, and the tests themselves skew toward exactly the workloads its customers care about.
Bottom line: the benchmark story is credible directionally, but provisional. Treat “state of the art” as Exa’s description of its own product until independent evaluators weigh in.
The Engine Room: Exa’s Largest Embedding Model Yet
Behind the benchmark scores sits the actual technical change. Exa says the gains in Exa 3.0 come from upgrades across its retrieval stack — and the centerpiece is the largest embedding model the company has ever trained, fed with more compute and more data than any previous version.
If “embedding model” sounds like jargon, here is the plain-English version. Search engines do not match keywords the way a Ctrl+F search does. They convert your query and every page on the web into long lists of numbers — embeddings — that capture meaning, and then they find pages whose numbers sit closest to the query’s numbers. A bigger, better-trained embedding model understands meaning more precisely: it knows that “cheap flights next Friday” and “budget airfare this weekend” are asking the same thing, and it can tell a coding question about Python decorators from a home-decor question that happens to share a word.
For AI agents, embedding quality is the whole game. An agent fires off dozens or hundreds of queries in a single task, and every irrelevant result it has to read costs tokens, time and money. If the first ten results are the right ten results, the agent finishes faster and cheaper. That is why Exa frames retrieval quality and cost as two sides of the same product — and why the embedding upgrade and the new cheaper tier launched together in Exa 3.0.
Exa also says its search latency is configurable between 450 milliseconds and 1 second, which matters for agent loops where the model is waiting on search results before it can take its next step. A slow search API turns an agent into a patient agent; a fast one keeps the reasoning loop tight.
Dynamic Highlights: 95% Fewer Tokens, Better Answers
The most practically exciting part of Exa 3.0 is not a benchmark at all. It is a feature called Dynamic Highlights, which is now generally available — and it attacks the single most annoying cost in agent search: token bloat.
Here is the problem it solves. When an agent searches the web today, the typical pipeline grabs entire pages — navigation menus, footers, cookie banners, sidebars, ads and all — and stuffs them into the model’s context window. Most of that text is noise. The model pays for every token it reads, so a five-page research task can burn through tens of thousands of tokens of boilerplate before the agent has learned a single useful fact.
Dynamic Highlights changes the pipeline: instead of handing the agent the full page text, Exa 3.0 extracts only the passage actually relevant to the query — the highlighted snippet, essentially — and returns that. According to Exa, this uses 95% fewer tokens than sending the full page while improving answer quality, because the model is no longer distracted by irrelevant content. Think of it as the difference between handing someone a highlighted paragraph and handing them the entire book.

For developers, the math is compelling. If search is already your agent’s biggest variable cost after the model itself — and for research-heavy agents, it often is — a 95% token cut on page reads can dwarf the savings from a cheaper per-request price. This is also the feature that most clearly shows Exa’s philosophy: search for agents is an infrastructure cost problem, and every token you do not send is money you do not spend.
One honest note: “95% fewer tokens” is Exa’s own measurement, and real-world savings will vary with the query. A page where the answer genuinely spans the whole document will not compress as neatly as a page with one relevant paragraph. But the direction is right, and it is included at no additional charge in both the Lite and Auto tiers.
Exa Lite vs Exa Auto: The Pricing Breakdown
Exa 3.0 introduces a two-tier search lineup, and the pricing is refreshingly simple — pay-as-you-go, no subscription, no minimum spend. Here is the full breakdown in USD:
| Exa Lite | Exa Auto | |
|---|---|---|
| Price | $4 per 1,000 requests | $7 per 1,000 requests |
| Best for | High-volume, cost-sensitive agent workloads | Maximum retrieval quality |
| How it works | Leaner retrieval pipeline, fewer compute cycles per query | Spends more compute on candidate generation, highlights and ranking |
| Results per request | Up to 10 results included | |
| Dynamic Highlights | Included at no extra charge | |
| Expanded index data | Included at no extra charge | |
A few more numbers that matter:
- Page contents: $1 per 1,000 pages, if you need full text beyond highlights.
- Free to try: new accounts get $20 in free credits — roughly 2,800 searches — and the free tier adds $10 every month.
- Deep search, Monitors and Exa Agent endpoints follow the same rate structure, and enterprise plans add volume discounts with custom rate limits.
- Rate limits: 10 search queries per second on the free plan, up to 25 on pay-as-you-go, custom for enterprise.
How should you think about Lite vs Auto? Exa’s own agentic-search chart plots the tradeoff: Lite scores around 79 on their internal measure at roughly $0.045 per task, while Auto scores about 80.5 at roughly $0.065. Those are per-task costs, not the published per-request rates, and the gap will differ for every application — but the shape is clear. Auto’s extra compute buys a modest quality bump, not a revolution. For most high-volume workloads — a support agent answering thousands of questions, a monitoring agent scanning the web on a schedule — Lite is the sensible default, and you only reach for Auto when retrieval quality is the bottleneck you can actually measure.
One more pricing reality check: at $4 per 1,000 requests, each search costs four-tenths of a cent. An agent that runs 10,000 searches a month costs $40 on Lite. That is genuinely cheap infrastructure — cheaper than a single coffee-shop latte per thousand searches, to put it in human terms. It is also why Exa can plausibly pitch itself as the default search layer for agents: at these prices, the switching cost of trying it is basically zero, especially with $20 of free credits on the table.
A Bigger, Fresher Index: What’s New in the Data
The second pillar of Exa 3.0 is data — billions of new documents, deliberately chosen for the things agents get asked about. Exa says it has seen an explosion in usage from personal AI assistants, so it prioritized expanding the index for those use cases. The new coverage includes:
- 1.27 billion ecommerce listings with prices and availability — the raw material for shopping agents and price-tracking bots.
- 211 million+ local businesses with hours and ratings — for “find me a plumber open now” style queries.
- Real-time news, sports, weather and market data — the freshness layer that static indexes struggle with.
- Flights, local events, transit and podcasts — structured, time-sensitive data that agents constantly need.
- Plus the existing coverage of people, companies, code and publications.

This list reads like a map of the agent economy’s most common questions. A travel agent needs flights and events. A shopping agent needs listings with live prices. A research agent needs publications. Exa is not trying to index “the whole web” in the abstract Google sense; it is indexing the kinds of web pages agents actually read, and growing the index around customer demand. That focus is arguably the smartest strategic move in the whole 3.0 release — because an index that is deep in the right places beats an index that is shallow everywhere.
The ecommerce angle deserves special attention. Product listings with prices and availability are exactly the data that retailers are starting to fence off from agents — as we covered in our piece on why major websites are blocking AI agents, companies like Yelp now require data licensing deals and Delta blocks agent bookings outright. An independent index with 1.27 billion listings is Exa’s answer to that trend: if agents cannot scrape every site freely, a licensed, structured index becomes the legal on-ramp.
Exa 3.0 vs the Competition
Exa 3.0 does not compete with Google. It competes with the other APIs fighting to become the search layer for AI products: Perplexity’s API, Brave Search, Tavily and Parallel. Here is how they line up on the factors that matter to developers:
| Exa 3.0 | Brave Search API | Tavily | Perplexity API | |
|---|---|---|---|---|
| Entry price | $4 / 1K requests (Lite) | $5 / 1K requests | Varies by plan | Varies by model |
| Index | Own crawler, 100B pages | Own index | Aggregated | Own + partners |
| Adapter focus | Built for agents from day one | General search API | Built for LLM apps | Answer engine first |
| Token efficiency | Dynamic Highlights (95% fewer tokens) | Standard snippets | Full page text bundled in results | Synthesized answers |
| Structured data | Ecommerce, local, flights, events | General web | General web | General web |
| Latency | 450ms–1s configurable | Fast | Fast | Varies |
A few honest distinctions are worth spelling out. Tavily’s approach — bundling full page text into every result — is convenient for simple apps but expensive at scale, which is exactly the cost problem Dynamic Highlights attacks. Brave scores well on citation precision in third-party comparisons, while Exa scores higher on groundedness. Perplexity’s strength is the consumer answer experience, not the raw API. And Parallel, the newest name on the list, is the one to watch as the direct “search for agents” challenger.
The fair summary: Exa 3.0 is the price leader with the most explicit agent-first design, but the gaps are not enormous, and none of these vendors has won yet. If you are already on Tavily or Brave and happy, Exa 3.0 is a reason to run a bake-off, not a reason to migrate on launch day.
Who’s Actually Using Exa?
Benchmark claims are cheap; customer logos are not. The most telling name in Exa’s customer story is Mozilla Firefox, which uses Exa to power Smart Window on desktop and Quick Answers on iOS — after evaluating multilingual quality, latency, safety and privacy, according to Exa. When a browser with hundreds of millions of users picks your search API for its AI features, that is a stronger endorsement than any benchmark table.

Beyond Firefox, Exa reports an explosion in usage from personal AI assistants — the consumer copilots that answer everyday questions and need fast, cheap, grounded web access. The Forbes Agentic 20 nod adds institutional credibility. What you will not find, at least publicly, is a roster of named enterprise deployments at the scale of, say, a Fortune 500 rollout. Exa is still a 51–200-person startup, and its customer proof is weighted toward product and developer adoption rather than big-logo contracts. That is normal for this stage — just calibrate your expectations accordingly.
On the trust side, the basics check out: Exa is SOC 2 Type II audited and honors GDPR and CCPA data-subject rights. Enterprise plans offer zero data retention, HIPAA support with a BAA, SSO and SCIM. For teams handling sensitive data, those are the checkboxes that matter, and they are checked.
Why AI Agents Need Their Own Search Engine
To understand why Exa 3.0 exists at all, zoom out to what is happening to the web in 2026. AI agents are everywhere — coding assistants that look up documentation mid-task, shopping copilots that compare prices, research agents that read dozens of sources before answering. All of them need to search the web constantly, and the old search stack was never designed for them.
Traditional search engines optimize for a human who will read a results page. They return ten links and let the person do the judging. An agent cannot “skim” — it reads everything it is given, pays for every token, and needs structured, citable facts it can reason over. Give an agent a human search page and you get the worst of both worlds: slow, expensive, and full of markup meant for eyeballs.
That mismatch is creating real friction across the industry. Just this week we reported on Yelp, Delta, JPMorgan and Charles Schwab actively blocking AI agents — some over data licensing, some over security and liability. At the same time, the orchestration layer is exploding: Anthropic’s dynamic workflows can now spin up to 1,000 subagents per run, and models like Inception Mercury 3 are pushing past 1,000 tokens per second. Faster agents running in larger swarms means more searches per task, not fewer — which makes the per-search cost and token efficiency that Exa 3.0 targets the exact right battle to fight.
The tension is the story of the agentic web in one frame: sites are walling agents out while agent infrastructure races to make them cheaper and faster. Exa 3.0 sits squarely on the infrastructure side of that fight — betting that licensed, structured, agent-priced search is how the web stays usable for machines.
Getting Started With Exa 3.0
Trying Exa 3.0 is deliberately low-friction. Sign up in the API dashboard and you get $20 in free credits — about 2,800 searches — with no subscription and no minimum spend, plus $10 added to the free tier every month. The API is documented on Exa’s developer docs, and the whole thing is pay-as-you-go from there.

A sensible first experiment: take one agent workflow you already run — a research task, a monitoring job, a support bot — and run it twice, once against your current search setup and once against Exa Lite with Dynamic Highlights enabled. Measure two things: the quality of the final answers (did the agent get the facts right?) and the token bill (how much context did the search step consume?). That A/B test tells you more than any benchmark table, because it measures your workload, not Exa’s.
If you want the story straight from the source, Exa is hosting a live Q&A with its Head of Product, Shyam Kumar, and Head of R&D, Liam Hinzman, on Wednesday, October 21 at 1pm PT. Launch-week Q&As are usually where the awkward questions — “why did you pick these benchmarks?”, “what does Lite actually cut?” — get real answers, so it is worth attending if you are evaluating seriously.
The Honest Verdict
So: is Exa 3.0 the real deal? Here is our balanced take.
What is genuinely impressive. The product thinking is excellent. Exa 3.0 attacks the two costs that actually dominate agent search — per-request price and per-page token bloat — at the same time, and the $4 Lite tier plus 95%-fewer-token highlights is a combination no competitor currently matches on paper. The expanded index is strategically smart: ecommerce listings, local businesses and real-time data are exactly what agents ask for. And the Firefox deployment is real-world validation that matters.
What needs a discount applied. The “#1 on 8 benchmarks” headline is a company-run claim without published raw outputs or independent reproduction. The leads over rivals are real but thin — fractions of a point on some tests. Treat the rankings as directional until third parties confirm them.
Who should try it. If you are building or running AI agents that search the web at volume — research bots, coding assistants, monitoring agents, shopping copilots — the $20 in free credits makes a trial essentially free, and Lite’s pricing is the cheapest credible option in the category. Run your own A/B test and let your workload decide.
Who should wait. If you need full page text bundled into every result by default, or you are already happy on Tavily or Brave at your current volumes, there is no emergency. Exa 3.0 is a strong new option, not a mandatory migration. And if your use case is consumer-facing answer quality rather than agent infrastructure, Perplexity’s answer engine remains the more natural fit.
The bigger picture is what makes Exa 3.0 worth watching regardless. Search is being unbundled: one stack for humans, another for the software that increasingly searches on our behalf. Exa 3.0 is the clearest bet yet that the agent stack becomes its own multi-billion-dollar infrastructure layer — priced per thousand requests, measured in tokens saved, and judged on benchmarks most people have never heard of. Whether Exa wins that layer or not, the layer itself is now undeniable.
Exa 3.0 FAQ
What is Exa 3.0?
Exa 3.0 is the October 2026 release of Exa Labs’ web search API, built specifically for AI agents rather than human searchers. It adds a larger embedding model, a cheaper $4-per-1,000-requests Lite tier, Dynamic Highlights that cut token use by 95%, and billions of new indexed documents.
How much does Exa 3.0 cost?
Exa 3.0 is pay-as-you-go with no subscription: Exa Lite costs $4 per 1,000 requests and Exa Auto costs $7 per 1,000 requests, each including up to 10 results per request. Page contents cost $1 per 1,000 pages. New accounts get $20 in free credits (about 2,800 searches), and the free tier adds $10 every month.
Is Exa 3.0 really better than Perplexity and Brave Search?
Exa claims first place on all eight benchmarks it ran against Perplexity, Brave Search and Parallel — including 91.8% on BrowseComp and 88.2% on DSQA. But the tests were run by Exa itself with no published raw outputs, so treat the rankings as a company claim until independent benchmarks confirm them. The leads on some tests are fractions of a point.
What are Dynamic Highlights?
Dynamic Highlights is an Exa 3.0 feature, now generally available, that extracts only the passage relevant to your query instead of returning the full page text. Exa says it uses 95% fewer tokens than full pages while improving answer quality, and it is included at no extra charge in both tiers.
Who uses Exa?
Mozilla Firefox uses Exa to power Smart Window on desktop and Quick Answers on iOS. Exa also reports surging usage from personal AI assistants, and the company was named to the Forbes Agentic 20 list of leading agentic-AI companies.
How is Exa different from Google Search?
Google is built for humans: it returns links for a person to read. Exa 3.0 is an API built for software: it returns structured, citable results an AI agent can feed directly into its reasoning, with pricing and token efficiency designed for high-volume machine workloads. Exa also runs its own crawler and index rather than wrapping Google or Bing.
Sources: Exa’s official announcement at exa.ai/blog/exa-3 and the company’s AI-assistant use-case page at exa.ai/use-cases/ai-assistants.
