Exa’s Agent Ultra Sends A Swarm Of Subagents At Deep Research, And The Benchmarks Say It Works

Posted on by
Contents show
A flat vector illustration of one large white magnifying glass with six small thin-line magnifying glasses fanning outward and small amber dots scattered where they point, representing Exa Agent Ultra, the subagent swarm for deep research and entity enrichment

The hardest research tasks share a shape: they run to exhaustion, and Exa Agent Ultra is built around that shape rather than around chat, which is why its benchmark story is about completeness rather than eloquence. Every company in a market, every paper implementing a technique, every portfolio company that announced a leadership change this quarter. Frontier chat models handle these tasks poorly not because they lack intelligence but because exhaustion is a workload problem, thousands of sources, many domains at once, no natural stopping point short of completeness. Exa's answer, shipped 26 September 2026, is Exa Agent Ultra: the highest effort level of Exa Agent, which divides a task into subtasks, assigns subagents to research multiple domains in parallel, and coordinates results across what the company describes as thousands of sources. In effect, deep research becomes an API primitive: you describe the list, the system runs the swarm, and you receive entities with cited evidence at a metered price per task.

The claim on the tin is strong: Exa Agent Ultra outperforms Opus 5.5, GPT-6 Astra and Perplexity Agent on their maximum effort settings across four benchmarks, with lower cost per task. And the release's own framing points at who it is for: the hardest research, where completeness and accuracy matter most, not the everyday queries a search bar already handles. Claims of that shape deserve both attention and caution, because deep research benchmarks are young and the companies that build the tools also grade them, so this article presents the numbers alongside the construction details that determine how much weight each can carry. This article works from Exa's own announcement, read directly, and does what a benchmark-led release deserves: explains the architecture, presents the numbers with their grader's own caveats, and maps where this kind of API fits real workloads, from due diligence to the lead-generation and list-building tasks your marketing team runs.

How the swarm works

Exa Agent's architecture is a work decomposition system. A task arrives, the system divides it into subtasks, and assigns subagents to research multiple domains at the same time. Intelligence is allocated proportionally: frontier models go where a subtask needs them, faster cheaper models go where they are enough. Exa Agent Ultra, as the highest effort mode, is built for the cases where completeness and accuracy matter most: deep research, list-building and entity enrichment, any task that needs to run to exhaustion. The efficiency logic is the same one that governs single-agent work: most subtasks do not need the biggest model, and the ones that do are usually the judgment calls, like whether a company plausibly sells to hospitals, rather than the mechanical ones, like finding a funding page.

The practical design detail is list continuation: pass the rows you already have and Exa Agent Ultra excludes them from future results, which turns a one-shot list build into an incremental, ongoing enrichment process, and which is also how Exa Agent Ultra avoids re-charging you for knowledge you already hold. For a monitoring feed, that is the Exa Agent Ultra feature that makes the API a pipeline component rather than a toy, and it pairs naturally with the routing discipline this site covered in model routing: match the effort level to the task, and pay for exhaustiveness only where exhaustiveness pays. For a market map or a monitoring feed, that is the difference between a demonstration and an operational tool.

The examples Exa publishes for Exa Agent Ultra read like a product requirements document for three audiences, and they are worth listing because they define the task shapes the system is built for. Model providers assembling training data: every paper and repo implementing a technique, every benchmark with a leaderboard and its top submission, and verification of hard-to-check criteria like whether a model released weights rather than just an API. This is deep research where the cost of a wrong row is a corrupted dataset, which is why the citations matter as much as the recall. Financial services: European battery-recycling market maps with founders, funding and customers per company; KYC research across filings, press, court records and regulator sites, synthesized into one record with citations; portfolio monitoring for layoffs, CFO changes and facility announcements. These are regulated workflows, where an uncited claim is not just weak but unusable, and where entity enrichment is a compliance task wearing a spreadsheet's clothes. Go-to-market teams: account lists like every US company making browser-automation tooling, and enrichment fields requiring judgment, such as whether a company sells to hospitals, with a cited URL as evidence. For this audience the pitch is simple: better lists at lower cost than manual research, with the receipts attached to every row, and Exa Agent Ultra is priced to be compared against the salary hours it replaces rather than against other software, which is the comparison its buyers actually make. Whether that comparison lands in your favour is exactly what the audit stage measures before the list reaches a customer.

That last detail, evidence-backed fields with citations, is what separates this category from a spreadsheet of guesses, and it is why the benchmark design below matters: a research product for entity enrichment must be auditable, or its output is unusable in exactly the domains, due diligence, compliance, legal, where exhaustive lists are worth the most.

The benchmark results, with their construction

Exa evaluated Exa Agent Ultra on four benchmarks. The headline numbers, from the announcement:

  • WANDR: 12.6 percent above Opus 5.5, 103 percent above Perplexity, 213 percent above GPT-6 Astra, at half the cost per task of Opus 5.5. WANDR, Wide And Deep Research, tests whether a research agent can find a large list of qualifying entities, wide, with evidence-supported accuracy across many fields, deep, and tasks are modeled on entity-enrichment use cases such as due diligence and legal research.
  • DeepSearchQA: 4.7 percent above Perplexity, 10.1 percent above GPT-6 Astra, 21.0 percent above Opus 5.5, at 46 percent lower cost per task than GPT-6 Astra.
  • WideSearch: 5.2 percent above Perplexity, 7.7 percent above GPT-6 Astra, 14.1 percent above Opus 5.5, at the lowest cost per task of the four.
  • Find-All Company: 1,579 percent above Opus 5.5, 2,069 percent above GPT-6 Astra, 2,401 percent above Perplexity, at the lowest cost per entity found. The enormous percentages reflect a base-rate effect: the competitors found so few entities that small absolute gains produce huge relative ones, which is a completeness demonstration rather than a quality comparison.

The grader disclosure deserves attention because it is more transparent than most. Exa states its grader for Exa Agent Ultra shares the same vendored evaluation logic as the upstream WANDR repository, with the core differences being the contents tool, Exa itself, the transport logic, and a cheaper judge model, gpt-6-luna. Where competitors had published results on this grader harness, Exa reports the published figure; where none existed, Exa ran the benchmark itself. That means the Exa Agent Ultra comparison is partially self-played, and the WANDR cost-per-task advantage also benefits from Exa's own infrastructure serving the contents tool. The quality claim rests on the shared grader logic, which is the strongest part of the construction; the cost comparisons are less independent.

What the subagent-swarm pattern changes for marketing and research work

For the readers of this site who run marketing, research or lead-generation, the Exa Agent Ultra release marks a category shift worth naming: deep research has become an API primitive with a price per task, rather than a person's afternoon or a fragile in-house agent script.

The old ways of building a list were search queries, scraped directories and manual verification, in decreasing order of accuracy and increasing order of cost, and the reason deep research APIs displace them is that the verification step is built in rather than bolted on: every enriched field arrives with the URL that supports it, so the analyst's job shifts from finding to auditing. The new option is a task description: every company matching this criteria, enriched with these judgment fields, each backed by a cited source. Where deep research replaces human hours, the cost arithmetic is lopsided, and where it replaces an in-house agent stack, the comparison is maintenance versus a metered API, with the API winning on anything the team does not run at constant volume. Either way, the cited-URL evidence requirement is the feature that makes the output auditable, which is what makes it usable in due diligence or compliance contexts rather than only in growth work.

The pattern also generalises beyond Exa: purpose-built deep research agents with dedicated indexes, parallel subagents and effort tiers are becoming a product category alongside the general chat assistants, and the reason is architectural. Exa Agent Ultra is one instance of it, and its benchmark table is the evidence that the instance works. A system built for exhaustion, with its own index and parallel subagents, beats a general model prompted to pretend, and the benchmarks here, whatever their construction caveats, are consistent with that structural argument. The same division-of-labour logic appears in this site's coverage of agent harness costs and multi-agent setups: putting the right amount of intelligence at each step, and scaling the number of workers to the shape of the task, is where the economics live.

Building a list pipeline around it

For a team adopting this class of API, the architecture that works is a pipeline with Exa Agent Ultra as one stage, not the whole machine. The stages that matter: define the target criteria precisely enough to be checkable, because vague criteria produce vague lists at any price; run the API task and receive the candidate list with citations; apply your own verification to a sample, checking whether the cited URLs actually support the enriched fields; and only then feed the list into outreach, CRM or analysis.

The verification stage is where trust in Exa Agent Ultra is built or lost, and the citation design makes it cheap. A sample audit of fifty rows tells you the error rate of Exa Agent Ultra on this task class, and if the error rate holds on the second audit, the pipeline can scale with periodic spot checks rather than full review. Teams that skip the audit stage are buying answers without calibration, and the failure mode is silent: an enriched field that is confidently wrong looks identical to one that is right, until it reaches a customer.

The second design decision is cost control. Exhaustive deep research is metered per task, and the difference between a well-scoped task and a sprawling one is measured in dollars as well as quality. The discipline is the same one this site recommends for agent spending generally, from inference caching onward: measure cost per completed task, cap the task's scope, and expand by increments using the exclusion mechanism rather than by re-running from scratch. A list pipeline built this way is auditable, budgetable and repeatable, which is what turns a research API into infrastructure.

Where the numbers stop

  • All benchmark results for Exa Agent Ultra are Exa's own evaluations on its own infrastructure, with the WANDR grader sharing vendored logic with the upstream repository but using Exa's contents tool and a third-party judge model. No fully independent replication existed at publication.
  • Where competitors had published results on this grader, Exa reports the published figures; where none existed, Exa ran them itself. The mix means the table blends published and self-run numbers, and the announcement does not itemise which is which.
  • The Find-All Company percentages for Exa Agent Ultra are relative improvements against very low competitor baselines. The absolute entity counts, which would show how many each system found, are not published in the announcement, so the completeness gap's true size cannot be computed from the release alone.
  • Cost-per-task comparisons favour Exa's architecture partly because Exa serves the contents tool the agents consume. That is a legitimate vertical-integration advantage for Exa Agent Ultra, and a real one for customers, but it is not a like-for-like comparison of reasoning quality at equal information access.
  • The four benchmarks measure wide-and-deep entity research. General reasoning, writing and conversational quality are out of scope, and nothing here suggests Exa Agent Ultra replaces a frontier chat model for those tasks.
  • I have not run the API. Every figure here is cited from Exa's release.

FAQ

What is Exa Agent Ultra and how does its subagent swarm work?

It is the highest effort level of Exa Agent, an API for deep research released by Exa on 26 September 2026, and Exa Agent Ultra is its most thorough configuration. The name describes the work: Exa Agent Ultra coordinates many agents across thousands of sources, and its target workloads are the ones where stopping early is the failure mode, which is what distinguishes deep research from ordinary search. The system divides a research task into subtasks and assigns subagents to research multiple domains in parallel, coordinating across thousands of sources, with frontier intelligence allocated where subtasks need it and faster models where they suffice. It is built for tasks that must run to exhaustion: deep research, list-building and entity enrichment, with results backed by cited URLs.

How does Agent Ultra compare to frontier models?

On Exa's published benchmarks, it outperforms Opus 5.5, GPT-6 Astra and Perplexity Agent on their maximum effort settings: 12.6 percent above Opus 5.5 on the WANDR wide-and-deep research benchmark at half the cost per task, with smaller single-digit leads on DeepSearchQA and WideSearch, and a very large relative lead on Find-All Company, a completeness benchmark where competitors found few qualifying entities. All figures are Exa's own evaluations using a grader that shares logic with the upstream WANDR repository but runs Exa's own contents tool, so independent verification is still pending.

Who should use a deep research API like this?

Teams whose work is exhaustive list-building and entity enrichment: market mapping for due diligence, KYC research synthesized across filings and court records with citations, competitive monitoring, and go-to-market list building with judgment-based enrichment fields like does this company sell to hospitals, each field backed by a cited source. The API is priced per task and supports incremental expansion by excluding rows you already have, which suits ongoing monitoring as well as one-shot builds. Teams whose research is conversational or analytical rather than exhaustive are better served by general models, and Exa Agent Ultra makes no claim in that direction.

The short version

Exa Agent Ultra is a subagent-swarm deep research API that divides tasks across parallel workers with intelligence allocated per subtask, built for entity enrichment and list-building at exhaustion, and Exa's benchmarks show it beating Opus 5.5, GPT-6 Astra and Perplexity Agent at their maximum settings on wide-and-deep research tasks, at lower cost per task. The architecture is the story: exhaustion is a workload problem, and a system built for it, with a dedicated index, parallel subagents and cited evidence, structurally beats a general model prompted to do the same job. The benchmark construction favours Exa, the grader runs Exa's own tools, and the Find-All percentages are base-rate effects, so the numbers deserve the caveats they got. But the category, research-as-an-API with a price per task, is real, and for list-building, due diligence and lead enrichment it changes the cost arithmetic from human hours to metered calls.

Sources
Subscription Form