Exa Release Notes

Today we announced ATLAS, a new benchmark that grades the performance of search agents across 547 wide search tasks. At this point, with WideSearch, DeepSearchQA, BrowseComp, WANDR, and many others, it is fair to ask: does the world really need another search benchmark? We think it does. Current benchmarks are either substantially memorized by frontier models, fail to reflect real-world search tasks, or are not precisely graded or reproducible.

Designing an effective search benchmark is a tough challenge. It needs to test agents on tasks of sufficient scale and difficulty while still being able to fairly grade them on completeness and accuracy at a manageable cost. The current crop of search benchmarks all make significant compromises, either in the tasks themselves or how they’re graded, that prevent them from achieving these goals.

We found that the best search agents of today are highly accurate; more than 90% of the entities they find are correct. However, they fail to find a third or more of possible entities. ATLAS is designed to help us track and ultimately close that completeness gap.

What is a search benchmark trying to measure?

Agentic research is used for difficult and high-value tasks. Either each individual entity represents real economic value, or the risk associated with missing a single critical entity is high. In other words, getting everything correct matters. For example:

• In a search for companies matching complex criteria, returned entities are often sales leads. Say there are 743 companies matching the search criteria but a search agent only finds 412. If you can close 4.3% of leads at an ACV of $85K, those missing 331 leads are worth about $1.2M, and that’s just one search.
• Investment firms may research dozens of companies before spending millions on a purchase. Due diligence for these deals requires exhaustive search for adverse media or for any prior art that could affect patents. Finding hundreds of documents but missing just one piece of vital information can put the entire investment at risk.

When research tasks are used to support agentic workflows, the ultimate goal is to reach a level of confidence in an assessment that justifies an action: a campaign, an investment, a partnership, etc. For that to happen, a result needs two qualities:

• Accuracy: entities and their properties must be correct and the information must be attributable to a trusted source.
• Completeness: All available entities must be found and gaps in available information must be known.

What does a useful benchmark look like?

To be able to successfully and fairly grade agents on completeness and accuracy in wide research tasks, a benchmark needs to meet four criteria.

1. Tests search, not memory

Tasks must require information the models don’t already know. If frontier models can answer from memory, the benchmark is not useful. In practice, given that frontier model training cycles have historically lasted 3-5 months,[1] tasks much more than six months old lose value quickly.

2. Difficult

Tasks have to be hard and at least some need to be large. If tasks are too easy, the benchmark will quickly become exhausted and models will cluster together. A benchmark is only useful if it separates out competitors.

3. Based in real demand

Tasks have to resemble actual work users want research agents to perform. Ideally they should be grounded in real queries. If every task in a set is made up by a research team, a high score might not reflect real-world results.

4. Economically viable

For a benchmark to succeed at scale in the long term the cost of creating tasks needs to be manageable, as does the cost to an assessor to run the benchmark against the models they want to test. The more a benchmark costs to run, the less it will be used.

The big problem: grading

Most wide search benchmarks begin with a similar format:

1. Give the agents a list of tasks to complete. The form is: find all entities that match X conditions, and enrich Y attributes.
2. Grade the performance of the agent for accuracy and completeness. Rank agents by how well they perform the task, as well as how long they take, and how much it costs.

Step one is straightforward. It’s necessary to regularly refresh tasks to prevent memorization and maintain the benchmark’s value. Research teams can create tasks, or source them from user surveys. Alternately, tasks can be built from real topics sourced from an agent provider. This second approach allows for a continuous stream of tasks that are known to reflect real user demand.

The second step is trickier. We’ve already established that the research tasks have to be difficult, so how can the answers be confidently judged? There are two main approaches to this problem. One approach is to build a “golden” answer key with complete and accurate answers. The second is to instead create a rubric to judge answer quality and have an LLM judge grade each response at run time.

Traditionally, each approach has required some major compromises.

Golden Answer Key

The big problem with an answer key is how to make an accurate and complete answer key for a sufficiently difficult task. Benchmarks built on answer keys usually use teams of human researchers to compile the key. This can get you an accurate and complete answer key but there are limitations.

First, creating the answer key is too expensive to allow for regular refreshing. The most common benchmarks using human derived answer keys are now 9-18 months old[2]. That’s more than enough time for the data to be built into a model’s weights[3]. Second, the need to conduct intensive human research limits the practical size and difficulty of the benchmark.

This means that for golden key benchmarks like DeepSearchQA and BrowseComp, all search agents perform reasonably well and there is relatively little separation between the best and worst performers[4].

In short, benchmarks that use answer keys can test completeness and accuracy of tasks, but they struggle to produce enough difficult tasks and to stay fresh enough to be useful.

LLM Judge

In this type of benchmark, a second LLM grades answers on whether the cited source supports the model’s claim for each entity and cell. The task has a “target” number of entities to retrieve, rather than working from a complete list.

The key advantage of LLM judges is that they don’t require the answer to be known ahead of time. They can test on larger tasks and can gracefully handle facts that change over time, like new hires for key positions in a company search. However, they introduce several of their own critical limitations.

First, they don’t test for completeness. If a model finds the target number of records, it can score full marks, even though in many real-world research tasks, missing one important entity can be critical. Second, they grade on whether the quoted source supports the agent’s answers, which is not the same as whether the answers are true. Finally, the need to run the LLM judge against each agent’s response makes these benchmarks orders of magnitude more expensive to run and grade.

In other words, LLM judges can help overcome the freshness and difficulty issues faced by golden key-based benchmarks, but they can’t reliably track progress towards completeness and accuracy in research tasks.

Exa’s approach to building a better benchmark

When designing our benchmark, Exa was not satisfied with either approach. Ultimately, we decided that the limitations of LLM judges in scoring accuracy and completeness were not acceptable and that an answer key was necessary.

The challenge then was to build an answer key that could be refreshed regularly enough to prevent memorization and that covered more and harder tasks than current answer key benchmarks. In the end, that meant building an automated pipeline that would allow agents, rather than human researchers, to build a task list and a comprehensive answer key.

For an in-depth look at how we built out ATLAS, read the technical paper. Here’s the short version:

Tasks start from topics our customers search for heavily. We use aggregate topic names and never read raw queries. For each task, two builder agents, one on ChatGPT and one on Claude, independently hunt for every entity that fits. Both search with a shared pool of Exa, Brave, and Perplexity results with the provider names hidden, so neither can lean on any one index. A separate verifier, using a search index outside that pool, checks every row against its evidence. Each column has to clear three gates before it’s kept: its values must come from pages beyond the ones that found the entity, a model can’t recite them from memory, and no single page answers the whole column. Only about one in five proposed columns makes it.

Then we tried to break the key. Twelve research systems ran every task, and every disagreement with the key became a lead. Judges from two model families researched each dispute without knowing which system gave the answer, and the key changed only when both agreed on verified evidence. Anything nobody could settle is left ungraded rather than guessed. Building and checking one task took a median of 8 agent-hours and about 1,200 searches, close to 4,800 agent-hours across all 547 tasks. A sampled audit puts wrong gold values under 1%. That cost is paid once. The result is an answer key that represents the best information a very well-resourced agent could be expected to know, and that takes into account information gaps and uncertainty.

After that, grading is a table comparison. While an LLM is still used for fuzzy matching, 95.5% of cells are scored without it, and grading one system on the whole benchmark costs about 24 cents. Because the pipeline is automated, we can rerun it on fresh topics every three to six months and stay ahead of what models have memorized.

Where ATLAS succeeds

Looking back at our success criteria for a benchmark, let’s judge how well ATLAS succeeds.

1. Tests search not memory

In a closed-book run (i.e. with no access to search tools), agents return many correct answers on existing benchmarks. On ATLAS, agents fail without web search.

2. Difficult

Exa Agent Ultra, an Exa model specially designed for intense research, gets all attributes for each entity correct about two thirds of the time. Of the frontier models, Opus 5.5 does best at a little under 50 percent, leaving plenty of headroom before the benchmark is exhausted.

3. Based in real demand

All 547 tasks in the current version of ATLAS are sourced from real search topics, with the automated pipeline ready to refresh with a new set of tasks when needed.

4. Economically viable

Using agents to create the answer key allows the benchmark to be built and refreshed at a lower cost than human-generated answer keys. Meanwhile, the cost per graded answer to run the benchmark comes in at 0.04 cents, compared to 88 cents for the LLM-judged WANDR benchmark when we ran it.

What does this mean for you?

We built ATLAS primarily to tell us how we were tracking towards our goal of a research agent that you can trust to power critical decision-making workflows. This requires the ability to conduct large and difficult research tasks comprehensively and accurately.

ATLAS is the most capable benchmark yet at balancing the need to test on a large set of difficult tasks, and confidently grade for the completeness and accuracy of an agent’s responses. That means that a high score on ATLAS is stronger evidence for real-world effectiveness than any other benchmark provides.

A few takeaways from our runs of ATLAS so far:

• There is still a lot of progress to be made on comprehensive search at a reasonable cost. Max-effort search agents like Exa Agent Ultra performed better than cheaper counterparts, but no agent run costing less than $1 per task achieved a success rate of over 50%.
• Even the most expensive search agents miss about 1/3 of the golden results, indicating substantial room for web search to improve performance in use cases where completeness is critical.
• When we tested just search APIs by holding the model fixed in a harness, we found that the choice of search API made a big difference in quality with scores ranging 16% across different search backends.

In short, when it comes to difficult research tasks, no agent can yet guarantee completeness of results, but the model and search backend you choose make a significant difference in the amount and quality of results you will find.

What’s next?

While we are confident that the ATLAS answer key is of exceptional quality, we’re currently conducting a blind human audit of a sample of tasks to ensure that they meet our standard of 99% of gold values proving correct. We’ll also be open sourcing this benchmark in a few weeks so that you can run it for yourself against any agents you’re considering for challenging search tasks.

Cheers,
Kale Bogdanovs

Today we're adding 211 million places to Exa's index. Your AI agents can now search restaurants, museums, parks, shops, and local businesses around the world, with structured data including opening hours, ratings, and phone numbers.

AI agents and personal assistants need up-to-date business data to help users find the right places to go and take action in the real world. Exa Places is available to use today across all our endpoints at no additional charge.

Data your agents can use

Exa Places includes restaurants and cafés, hotels, retailers, healthcare providers, professional services, cultural attractions, outdoor spaces, and more.

For example, user might query Exa Places for:

Ramen restaurants in the SF Mission open Friday at 7pm with an upscale vibe

Museums in Mexico City with reviews mentioning activities for children

Independent bookstores with great review scores in Brooklyn

Exa understands the geolocation in your query and returns relevant businesses in that area.

Results returned from the Exa Places index include structured data like:

• Names, categories, addresses, and coordinates
• Opening hours and operating status
• Phone numbers and websites
• Ratings

The structured fields are efficient for giving agents information they can work with directly. A sample response:

When accuracy matters

Personal assistants

People ask assistants about places: where to eat, which pharmacy is open, what hotel is near the stadium. With Exa Places, the assistant gets each place's status, hours, and phone number in the same search call it uses for the web. It can check if a place is open before recommending.

Go-to-market operations

When teams sell to local businesses, they need lists that are current, like every dental practice in a metro area, every independent gym in Texas, or all the physical therapy clinics in Phoenix. Maybe it can be done by hand, but it’s onerous and prone to error. With Exa Places, a search like these returns the businesses with phone numbers, websites, and more. The agent can use the data to kick off automated workflows. With Exa Agent Ultra, the list can return every single match.

Market research

Analysts sizing a market or choosing a site need accurate counts. How many coffee shops operate within a mile of a planned location? Which towns in the state have no urgent care clinic? It’s a painstaking effort if a person tries to do this with web search. Exa Places makes it fast and easy to get current counts of places matching virtually any criteria.

Higher quality results

Across 400 queries, we compared Exa Fast, Perplexity Search, and Parallel Advanced on finding businesses with correct metadata.

We measured F1, which balances precision and recall for businesses returned with correct metadata (opening hours, phone numbers, location).

Exa Fast achieved a 51% higher F1 score than Perplexity Search and a 140% higher F1 score than Parallel Advanced.

Quality matters for agents. Missing a relevant place limits the options an agent can consider. Returning the wrong address or hours can make an otherwise good recommendation unusable.

Start building

Exa Places is live in the search API for all users at no additional cost. Get started in the developer dashboard.

Cheers,
The Exa Team

Today we shipped Exa Agent Ultra, setting a new state of the art for comprehensive deep research. As the highest effort level for Exa Agent, it coordinates many agents across thousands of sources, delivering higher-quality results than frontier models.

Agent Ultra outperforms Opus 5.5, GPT-6 Astra, and Perplexity Agent on their respective maximum effort settings, across a variety of benchmarks:

Frontier deep research

Exa Agent Ultra outperforms all frontier models on their native harnesses and sets the state of the art on WANDR, scoring highest on quality and lowest on cost-per-task out of all competitors benchmarked.

• WANDR. +12.6% over Opus 5.5, +103% over Perplexity and +213% over GPT-6 Astra, at half the cost per task of Opus 5.5.
• DeepSearchQA. +4.7% over Perplexity, +10.1% over GPT-6 Astra and +21.0% over Opus 5.5, at 46% lower cost per task than GPT-6 Astra.
• WideSearch. +5.2% over Perplexity, +7.7% over GPT-6 Astra and +14.1% over Opus 5.5, at the lowest cost per task of the four.
• Find-All Company. +1579% over Opus 5.5, +2069% over GPT-6 Astra and +2401% over Perplexity, at the lowest cost per entity found.

We evaluated Agent Ultra with WANDR (Wide And Deep Research), a benchmark that tests research agents for their ability to find a large list of qualifying entities (wide) with evidence-supported accuracy across many fields (deep). Tasks are modeled on entity enrichment use cases in domains such as due diligence and legal research. Our grader shares the same vendored evaluation logic as the upstream repository; the core differences lie in the contents tool (Exa), the transport logic, and a better and cheaper judge model (gpt-6-luna). Where Perplexity, Anthropic or OpenAI had published a result obtained on this grader harness, we report the published figure; where none existed, we ran the benchmark ourselves. Faded points are each system's low setting (low effort for Opus 5.5 and GPT-6 Astra, Perplexity Agent on low-effort Astra, and Agent Auto for Exa).

Lists without limits

Exa Agent divides a task into subtasks, assigning subagents to research multiple domains at once. It puts frontier intelligence where a task needs it and faster models where they are enough. As the highest effort mode for Exa Agent, Ultra is built for the hardest research, where completeness and accuracy matter most: deep research, list-building, and entity enrichment - any task that needs to run to exhaustion.

Examples:

Model providers

• Assemble training data from the web, such as "every paper and code repo implementing a given technique" or "every public benchmark with a leaderboard and its top submission".
• Verify hard-to-check criteria about an entity like "has a reproducible eval harness", or "released weights, not just an API".

Financial services

• Create market maps for diligence like "every company that does battery recycling in Europe, adding founders, funding, and customers for each".
• Research a customer for KYC compliance across many sources such as filings, press, court records, and regulator sites, synthesized into one record with citations.
• Monitor for signals, like "every portfolio company that announced a layoff, CFO change, or new facility this quarter".

Go-to-market teams

• Build account lists like "every US company that makes browser-automation tooling" or "every Series A-C fintech with a Head of Risk hired in the last 12 months".
• Enrich a list with new fields requiring judgment, such as "does this company sell to hospitals?", "who owns the data platform?", with a cited URL as evidence.

Ultra makes it easy to expand lists by passing the rows you already have, excluding them from future results.

Try Ultra today

Agent Ultra is available today. Read the docs here or try it out in our API Playground.

Cheers,

The Exa Team

Agent Ultra

Agent Ultra is Exa Agent's highest-effort mode, built for large list building, deep multi-source research, and criteria that are hard to verify. Set effort: "ultra" on a run.

Metered: billed at standard Agent usage rates, up to a default $20 per run. Set budget.maxCostDollars to change the cap.

Time budget: set budget.maxDurationSeconds (5 minutes to 3 hours) and the run wraps up with what it has found by then.

Stop early: POST /agent/runs/{id}/stop ends a run and keeps the results gathered so far.

Read the Agent Ultra guide →

There's growing demand to test models in more complex environments and on increasingly difficult tasks. One kind of test is temporal: with the benefit of hindsight, check whether a model could have found a solution that hadn't been published yet, or could have predicted what happened next.

Today we're launching Exa Snapshot, a capability that unlocks this type of testing, powered by over 400 billion webpage snapshots spanning 2 decades. Given a date, Exa returns results from the web as it was at that point in time.

Reasons to test with hindsight

Prevent web leakage and run reproducible evaluations

Let's say you're training an agent with reinforcement learning (RL) on tasks and answers written in June.

If the training run takes place in September, some of the solutions may already be on the web: in a paper, a GitHub pull request, or a blog post. Agents with search tools will find and submit the answer, and it's difficult to differentiate between a genuinely solved task and one copied from the web. Graders may even favor the second agent because copying requires less effort.

With Snapshot, you can set
snapshotAsOf
to a date in May. Search no longer returns the solutions, and researchers can trust that an agent with the correct answer solved it fairly.

In finance and quantitative trading

A similar problem shows up in finance. A backtest checks how a prediction model would have performed on a given day using only information available as of that day.

Quantitative researchers will often solve for this with point-in-time datasets that track stock price and other fundamentals - but nothing like this exists for the web. Trading signals derived from web page insights would take months of careful data collection to test.

With Snapshot, backtesting on a versioned web can be conducted immediately.

How to use it

You can use Snapshot by setting
snapshotAsOf
in our
/search
and
/contents
API endpoints (see the
docs
). As a reminder, this is still a research preview with ongoing development.

For full index coverage, rate limits, ZDR, and more, reach out to our team.

Try it out in the
dashboard
by setting a Snapshot date, or read the
docs
to get started.

Cheers,
The Exa Team

Exa Agent and Exa Connect in MCP

Exa Agent is now available within Exa MCP. Use it from Claude, Cursor, or any other MCP client when the task needs more than a single search call.

Enable the Agent tool with https://mcp.exa.ai/mcp?tools=agent_run, then call agent_run to run the agent through completion and return its output.

Exa Connect data sources are available through the Agent flow, so you can attach premium data partners when a run needs more than web search alone.

Read the Exa MCP guide → · Read the Exa Agent guide → · Announcement tweet →

Let's say your agent wants to know the current mayor of SF. The answer can be found in 9 tokens on a webpage from August 26th, but most agent search tools will pull 10,000+ tokens across many pages, bloating context windows and increasing LLM spend.

Today, we're excited to release a research preview of Dynamic Highlights: given a search query and set of URLs, a custom model dynamically selects just the right tokens across all of the content, reducing token count by an average of 95%, compared to full content of pages.

The Advent of Agent Search

Today, by volume, agents are the primary users of the web.

Agentic context consumption comes with tradeoffs - causing context rot, exhaustion of context windows and unreasonably high LLM costs. This is precisely why token efficiency has never been more critical to search and to Exa's mission.

To mitigate this issue, Exa highlights returned snippets of documents in which we returned spans of each document that were relevant to the agent dispatched query. While using highlights reduced search tokens by a factor of five, this still led to duplicate information: multiple URLs pulling the relevant tokens that answer a given query or not pulling enough information from highly relevant documents because of per document token budget constraints.

Dynamic Highlights

To solve these problems with highlights, we introduce Dynamic Highlights, a new kind of search primitive that views tokens, not documents, as the fundamental unit of information to retrieve. More explicitly, instead of extracting query-relevant tokens from each document independently, Dynamic Highlights chooses relevant tokens with awareness of all other returned documents.

Dynamic Highlights uses one forward pass rather than many parallel ones, concatenating all retrieved documents into a single unified document. This lets it dynamically allocate more tokens to better content and fewer (or none) to irrelevant documents.

Evals

Single Turn

For simple RAG where one search is allowed, we can dramatically reduce the number of tokens returned by search while increasing downstream eval performance.

The results below are from tasks spanning coding to general question-and-answer, averaged across four runs of exa auto with different character limits for dynamic highlights.

With a 12k character budget, relative to existing highlights dynamic highlights achieves a 40% average token efficiency gain with a notable quality increase of 3.8%

Agentic

Unlike single-turn search, agents run for longer when they don't find the answers that they are looking for.

Search results are a significant contributor to token budget in such cases. However, reasoning traces take up more than half of the token usage of the agentic searcher, depending on the difficulty of the task.

Therefore, every search token usage optimization must maintain a strict quality bar; otherwise, the increase in the number of searches will eat away at any token efficiency gains.

We use Exa Agent, effort Auto on BrowseComp and WideSearch, as well as Exa Agent, effort Medium for our internal company and people datasets. Although token efficiency varies significantly across different tasks, we can achieve a 30% reduction in agent token usage with a small quality gain (2.1% on average).

This 30% reduction directly translates into cost savings for our users and any search agent products built on top of our API.

How to use it

You can use dynamic highlights by setting dynamic to true in our API — see the docs. As a reminder, this is still a research preview with ongoing development but is our most advanced snippets solution.

Cheers,

The Exa Team

Dynamic Highlights (research preview)

Dynamic Highlights selects excerpts across the complete result set instead of treating each page independently. It gives more of the shared context budget to useful sources and gives less context to sources that only repeat information already returned.

Single-turn RAG: about 49% better token efficiency and 2.4% higher downstream quality with Exa Auto across coding and general QA evaluations.

Agents: about 30% fewer tokens across complete agent trajectories and 1% higher quality across BrowseComp, WideSearch, and internal company and people evaluations.

Requests that set dynamic: true require the Exa-Beta: dynamic-highlights-2026-08-28 header.

Read the Dynamic Highlights guide →

Today we're launching state of the art search over research papers and technical publications at Exa. We built a dedicated index of roughly 350 million publications and added around 30 million authors to our people index. Exa can now search across scientific literature using natural language, even when a query is vague, highly specific, or based on an imperfect memory of a paper.

Research paper search is available through the Exa API using the Publication category.

Find papers semantically (the way people actually remember them)

Researchers do not always search using a paper's exact title. Sometimes they remember a specific result:

In a German study of 145 cosmetic and personal care samples, what percentage contained NDMA and what was the maximum concentration found?

Other times, they remember only the broad outline:

Wasn't there a paper from the mid 70s looking for a threshold below which dietary cholesterol wouldn't harm arteries in primates, and they never found one? I think HDL dropped and LDL rose even though total cholesterol looked normal.

Traditional academic search works well when you already know the title, author, or exact keywords. It is much less useful when all you have is a result, an experimental setup, or a half-remembered description.

Exa searches over the meaning of the query and the contents of the paper. This makes it possible to retrieve a specific publication from factual clues or an incomplete recollection, without requiring the user to translate the question into the right academic keywords.

Evals

We created two new benchmarks reflecting common ways people search for research publications. The first tests known-item retrieval, where the searcher provides precise factual clues and the system must identify the paper they came from. The second tests tip-of-the-tongue retrieval, where the searcher describes a paper using vague, incomplete, or slightly incorrect details, closer to how someone recalls work encountered months or years ago.

Across this benchmark, Exa retrieved the correct paper for 82.8% of queries.

Exa had the highest recall and MRR of the systems tested while returning results in under 400 milliseconds on average.

We are also developing an evaluation for topical completeness. Instead of finding one specific paper, it measures how completely a search system recovers the important body of work around a topic, such as:

Floquet engineering methods for topological phase simulation in ultracold atoms

Finding a single known paper is useful. Understanding a field requires finding the surrounding work too.

How we built it

Research papers are unusually difficult documents to search. Important information may be buried in a long PDF, a table, an appendix, or a scanned document. The same paper may also appear at several URLs with different titles and metadata.

Our ingestion pipeline uses OCR and document-parsing models to turn research PDFs into high-quality searchable text. We combine this text with information about authors, institutions, publication histories, citations, and collaborators.

At query time, Exa searches both its web index and the dedicated publication index, then combines and reranks the results. This gives researchers the coverage of the open web with the structure of a purpose-built scientific index.

Try it

You can use it through the Search API today:

from
exa_py
import
Exa

exa
=
Exa
(
)

results
=
exa
.
search
(

"papers proposing alternatives to attention for long-sequence modeling"
,
category
=
"publication"
,
num_results
=
10

)

Try it in the Exa dashboard or read the Search API documentation.

Cheers,
The Exa Team

Publication research

We significantly expanded and improved research over academic publications.

• 350M publications: search across an index of 350 million publications.
• Richer organization and people results: searches now return both organizations and their affiliated people, each as a detailed, enriched profile spanning publications, top collaborators, research areas, and funding.
• Agentic people and organization search: agents can now search over people and organizations.
• Public retrieval benchmark: we released a public benchmark for publication retrieval.
• New publication search category: query scholarly results with category: "publication", which replaces the research paper category.
• Deprecated categories: the pdf, github, and tweet search categories are being deprecated.
• startCrawlDate / endCrawlDate: deprecated parameters are now ignored for all teams while remaining accepted for compatibility.

Query it via the API with the publication search category, or try it in the dashboard →.

Publication research

We significantly expanded and improved research over academic publications.

350M publications: search across an index of 350 million publications.

Richer organization and people results: searches now return both organizations and their affiliated people, each as a detailed, enriched profile spanning publications, top collaborators, research areas, and funding.

Agentic people and organization search: agents can now search over people and organizations.

Public retrieval benchmark: we released a public benchmark for publication retrieval.

New publication search category: query scholarly results with category: "publication", which replaces the research paper category.

Deprecated categories: the pdf, github, and tweet search categories are being deprecated.

startCrawlDate / endCrawlDate: deprecated parameters are now ignored for all teams while remaining accepted for compatibility.

Query it via the API with the publication search category, or try it in the dashboard →.

Exa Agent and Exa Connect in MCP

Exa Agent is now available within Exa MCP. Use it from Claude, Cursor, or any other MCP client when the task needs more than a single search call.

Enable the Agent tool with https://mcp.exa.ai/mcp?tools=agent_run, then call agent_run to run the agent through completion and return its output.

Exa Connect data sources are available through the Agent flow, so you can attach premium data partners when a run needs more than web search alone.

Read the Exa MCP guide → · Read the Exa Agent guide → · Announcement tweet →

We're releasing Exa Agent: a single API to access frontier web research capabilities at a fraction of the cost.

Exa Agent combines top language models with Exa's state-of-the-art web search tools to achieve the most exhaustive and accurate results now used in a variety of agentic products.

Exa Agent capabilities

Exa Agent is highly effective on deep research, list-building, and entity enrichment tasks.

When working with large datasets, Exa Agent divides the task into many subtasks and assigns subagents to research various domains at once. When researching, it uses a fusion of frontier and cost-effective models to find the most cost-effective methodology for the given research task.

To quantify performance, we used an array of public benchmarks that measure quality, latency, cost and token efficiency.

Our goal was to produce the best web research agent using methodologies like model fusion and Exa's token-efficient highlights model (which have shown up to 94% reductions in token usage). The result is significantly lower cost and latency for frontier performance.

Benchmark chart examples

WideSearch

BrowseComp

DeepSearchQA

FinanceAgent-V2

People-FindAll

Company-FindAll

WideSearch Methodology

WideSearch was introduced in August 2025 to evaluate agents' abilities to aggregate and structure atomic information about entities from across the web. The expected output is always a table, structured as a list of enriched entities.

The English version contains 100 tasks, with the number of required columns varying from 3 to 14. In our setup, we evaluate only the agent's final answer. We compute F1 over rows: a row is counted as successful only if the matched entity and all required enrichment columns are valid. Otherwise, it is counted as a miss.

We experimented with cell-level F1, but found it too permissive because it rewards isolated enrichment values even when the agent failed to ground them to the correct entity.

How teams are using Exa Agent

Companies building their own agents can plug into the Exa Agent API directly.

Finance Agents can take advantage of Exa Agent's ability to retrieve real-time data from across the web, aggregating it into any output format required.

Go-to-market agents can bring their own list of accounts or prospects to enrich, or use Exa Agent's list-building capabilities to generate lists of tens or hundreds of entities.

Agent output examples

Company Research

Data Enrichment

Code Review

Literature Review

Research all publicly available information on Databricks: recent funding, product launches, partnerships, executive hires, conference talks, and public GitHub activity.

Exa Agent API result

Run completed June 16, 2026

Compiled a compact Databricks company brief across funding, launches, partnerships, hires, events, and public GitHub activity. The main signal is momentum from a $134B February financing and a product narrative around governed enterprise AI agents, Lakebase, Genie, and OpenSharing.

Area Finding Evidence

1 Recent funding $7B announced in February 2026 at a $134B valuation

Databricks announced more than $7B total investment, including about $5B equity and about $2B additional debt capacity. Reports also cited discussions for a possible $165B-$175B round in June.

2 Product launches OpenSharing, Agent Bricks, Lakebase, Genie, AI Gateway, and Lakeflow

Recent launches cluster around governed enterprise agents, model access, serverless Postgres for AI agents, conversational analytics, open data/AI asset sharing, and data engineering.

3 Partnerships Frontier model, SI, open-source, and data ecosystem partnerships

The partner map includes Anthropic, Accenture, OpenAI, Linux Foundation OpenSharing, and ecosystem partners such as SAP, Stripe, Atlassian, Retool, LlamaIndex, Collibra, Deloitte, and Wipro.

• 4 more rows

Notes

Compiled from Databricks press releases, blogs, docs, GitHub pages, Reuters/MarketScreener reporting, Linux Foundation materials, and conference pages as of 2026-06-16. The June 2026 valuation item is reported discussion, not confirmed closed financing.

Try the API today

Exa Agent using auto effort will dynamically scale to the amount of compute required by the given task. Developers can also set an effort level for fixed costs:

Effort Cost Best for

minimal $0.012 / request Lightweight tasks, lowest cost

low $0.025 / request Simple lookups, narrow factual tasks, short answers

medium $0.10 / request Default starting point for most standard research tasks

high $0.50 / request Harder research, more citations, stricter completeness

xhigh $1.00 / request High-value tasks where completeness matters more than cost/latency

The API also allows developers to use structured outputs via the outputSchema parameter, and to bring their own data using input.data. Exa Agent uses many tools to operate efficiently over large datasets.

Exa Agent is available today. Read the docs or try it in the Exa API Playground: dashboard.exa.ai/playground/agent

Cheers,
The Exa Team

Introducing Exa Agent

We released a new class of frontier web research agents that are accessible via API.

Exa Agent API supports parameters including a natural-language query, effort mode, outputSchema for structured outputs, and input.data to build upon an existing dataset.

Read the Exa Agent API guide →

Introducing Exa Connect

Exa Connect gives Exa Agent live access to the world's public and private data. It launched with Similarweb, Fiber.ai, Baselayer, Financial Datasets, Affiliate.com, Particle, Jinko, and Additional Partners. You attach them via dataSources on POST /agent/runs.

Read the Exa Connect guide → · Announcement tweet →