Exa 3.0: What the Search API Claims on Eight Benchmarks and What You Can Check

Search APIs for AI agents are in the same phase LLMs were in last year: every vendor says it is number one, and the charts are drawn by the vendor. On October 19, 2026, Exa announced Exa 3.0, which it calls "the largest improvement to search quality we've ever released," and says it "ranks first on every benchmark we tested." That is a strong claim, so this post lays out what Exa actually published, what is a vendor claim, and how you can check it for your own workload.

TL;DR: Exa 3.0 claims versus checks

What did Exa 3.0 change?

Exa says the release rests on three improvements to the retrieval stack that help every search type:

• A larger embedding model, trained with more compute and data than any previous Exa model. Exa gives no dimensions or standalone benchmark for it.
• Dynamic Highlights, now generally available at no extra charge. A model picks the most useful tokens across all results, favoring the best pages and cutting duplicated content. Exa says it uses 95 percent fewer tokens than full page text.
• A bigger index. Exa says it runs its own crawler and tracks more than 2 trillion URLs while serving 100 billion pages. New data includes 1.27 billion e-commerce listings with prices and availability, over 211 million local businesses with hours and ratings, and real-time news, sports, weather and market data. Flights, local events, transit and podcasts are also named.

There is also a more efficient reranker for long documents. If you want the background on how Exa positions its index, our earlier piece on Exa Snapshot and versioned web search is a useful companion.

Search results with source citations, representing Exa web search for AI agents

What are the eight benchmark claims?

Exa publishes these 3.0 scores in its chart:

Exa states that it leads Perplexity, Brave and Parallel on all eight. The page we read does not list those rivals scores, does not give dataset sizes or scoring rules, and does not say what units the FindAll numbers use. A figure like 91.8% on BrowseComp is also not a measure of search alone: scores on agentic benchmarks depend on the model and agent harness wrapped around the search tool. That matters when you try to reproduce a headline number.

We found no independent measurement that adds to these figures.

What do Lite and Auto cost per task?

Exa also charts agentic search cost versus quality:

Exa says the two set a "new Pareto frontier for search" among the providers tested. The page does not name the other providers on that chart or describe the test. Note the small quality gap: 79 versus 80.5 for roughly 30 percent lower cost per task. In practice, a team running thousands of agent searches a day should compare cost per correct answer rather than cost per request, because a cheaper tier that needs a second query to reach the same answer erases the saving. If that gap holds on your tasks, Lite is the default to try first; if it does not, the cost advantage disappears.

Pricing changes are the part of a launch that hits budgets directly, so it is worth comparing with other agent-search offers we have covered: Cloudflare Web Search through AI Gateway, the Perplexity Search API and Parallel own web search eval. Each of them publishes its own evaluation too, which is exactly the problem.

Pricing tiers ladder suggesting Exa Lite and Auto search costs

Why vendor search benchmarks need care

Exa earlier comparison pages show the pattern. Its Brave comparison, tested between July 29 and August 2, 2026, reported higher people-search and company-search recall for Exa, while also showing Brave ahead on citation precision. That is a useful signal: a vendor that publishes where it loses is easier to trust than one that reports a clean sweep. The 3.0 announcement reports a clean sweep, without that detail.

Three habits help when reading any search benchmark:

1. Ask what was held constant. Same model, same harness, same prompts, same date? Web results drift daily.
2. Ask who picked the benchmarks. Eight benchmarks the vendor selected are not the same as a neutral suite. Our guide to OpenRouter web search benchmarks for agent grounding walks through what independent comparisons look like.
3. Ask about cost, not only score. A score gain that doubles tokens or calls may cost more per finished task.

How to test Exa 3.0 yourself

A practical test takes an afternoon:

1. Collect 50 to 100 real queries from your agent logs, including the hard ones.
2. Run each through Exa Lite, Exa Auto and your current provider, with the same downstream model.
3. Grade the final answers, not the raw links, using a rubric or a second model as judge, with a sample checked by hand.
4. Record total cost per correct answer, including tokens spent reading results. Dynamic Highlights should reduce this; measure whether it does.
5. Measure p50 and p95 latency from your own region.
6. Repeat a week later, because index freshness is part of the product.

Exa publishes comparison code in its public GitHub organization for earlier evaluations, which can serve as a template.

What this means for what you build or pay

If your agents search the web heavily, the practical facts are the $4 Lite tier, the free Dynamic Highlights, and the new structured data for products, local businesses and live feeds. If your agent needs flight, shopping or local lookups, the expanded index may remove the need for a separate API, but only if the data quality holds up in your tests. If you already use another provider, nothing in the announcement is enough evidence to migrate on its own.

A live Q&A with Exa Head of Product Shyam Kumar and Head of R&D Liam Hinzman is scheduled for Wednesday, October 21 at 1pm PT, which may answer the open methodology questions.

What does the pricing look like at scale?

The per-request prices are simple arithmetic, which makes them easy to sanity-check against your own traffic. At the list prices Exa published, a workload of one million search requests a month costs about $4,000 on Lite and about $7,000 on Auto, before any downstream model tokens. At 100,000 requests the same tiers come to $400 and $700.

The per-task figures on Exa chart are higher than a single request because an agentic task usually makes several searches and reads their results. About $0.045 per task for Lite implies a handful of calls plus the cost of reading them, though Exa does not say how many searches the test tasks used or which model drove the agent.

Two things follow for budgeting:

• Request price is only part of the bill. Reading results costs model tokens, and that is where Dynamic Highlights should help. A claimed 95 percent reduction versus full page text sounds large, but most pipelines already truncate pages, so your real saving will be smaller than that baseline suggests.
• Quality per dollar beats price per request. If Auto resolves a task in one search where Lite needs two, the cheaper tier can cost more per finished task. Count completed tasks, not calls.

Exa did not announce volume discounts or rate-limit changes in the page we read, so check the dashboard before planning a large migration.

What is still unknown

• Competitor scores and exact versions for all eight comparisons
• Dataset sizes, judging method and number of runs
• Whether the FindAll numbers are percentages or counts
• How much of the BrowseComp and DSQA result comes from the search tool versus the agent around it
• Latency numbers for Lite and Auto under load
• Independent reproductions

Bottom line

Exa 3.0 looks like a real upgrade on paper: a bigger index, cheaper Lite tier and token-saving highlights, with clear published prices. The first-place claim on eight benchmarks is the least verifiable part. Use it as a reason to run your own comparison, not as a reason to switch.

Related reading

• Exa Snapshot: versioned web search API
• Cloudflare Web Search API through AI Gateway
• Perplexity Search API and index debut
• Parallel web search eval for AI agents
• OpenRouter web search benchmarks for agent grounding
• Official: Introducing Exa 3.0 and Exa docs on evaluating search

Prices and scores are as published on October 19 to 20, 2026 and may change.