Exa's Websets V2 builds verified lists from the web 4x faster at about half the cost per match. Pay-as-you-go pricing replaces the subscription, and every user now gets API access.
- Exa released Websets V2 : 4x faster dataset building at roughly half the cost per match.
- Accuracy and recall are up 20%, and the gains apply automatically for all users with no code changes.
- The subscription is gone; pricing is now $0.10 per compute unit plus $5 per 1,000 searches.
- Every usage-plan user gets API access , billed from the same credits as Exa's Search API.
- On Exa's own benchmarks, V2 averaged $0.017 to $0.024 per verified match versus Claude Code, Gemini, and Parallel.
- Existing Starter and Pro users must opt in to the new pricing , and the switch can't be undone.
Exa has released Websets V2, an upgrade to its service for turning plain-English queries into structured datasets sourced from the web. Exa reports fourfold faster jobs, roughly half the previous cost, and 20% gains in both matching accuracy and recall. The company has also removed the monthly subscription for new users and opened API access to everyone. Existing integrations receive the performance improvements without code changes.
| Change | Exa's claim | Developer impact |
|---|---|---|
| Speed | 4x faster | Long-running research jobs finish sooner |
| Cost | About 50% lower | Usage is billed by compute and searches |
| Retrieval quality | 20% higher accuracy and recall | More eligible results with fewer incorrect rows |
| Access | No monthly subscription for new users | Dashboard and API access use shared Exa credits |
Lists that do not exist yet #
Websets targets set discovery: finding many entities that satisfy several conditions at once. Conventional search engines rank pages, while chatbots often produce short lists with uneven sourcing. Websets assembles a structured result set from live web data and Exa's index of more than 1 billion people, 70 million companies, 1.2 billion products, and 500 million research papers. Each row includes sources and an explanation of how it matches the query.
Queries can combine subject, geography, technical criteria, organizational details, and required evidence. Exa gives examples including:
- Research papers on protein design that use diffusion models and release their code
- Universities with labs researching solid-state batteries for electric vehicles, including the faculty leading each program
- Companies developing geothermal power plants with active projects in Europe
Websets uses language models to evaluate candidates against every condition in a query. This test-time compute can keep a job running for minutes or hours because the service searches, reads, reasons about, and verifies individual results. Cutting that runtime to one-quarter allows teams to run more searches and iterate faster on criteria.
Websets takes on agent loops #
Exa evaluated V2 on two internal benchmarks, Company FindAll and People FindAll. Each benchmark contains 20 complex queries, and a result counts only when it satisfies the stated criteria. The comparison included Websets, Claude Code, Gemini 3.1 Pro, and Parallel FindAll, reflecting the common alternative of repeatedly sending a general-purpose agent through search and verification steps.
| Benchmark | Queries | Verified matches | Matches per query | Cost per match |
|---|---|---|---|---|
| Company FindAll | 20 | 514 | 25.7 | $0.024 |
| People FindAll | 20 | 613 | 30.65 | $0.0169 |
Exa says Websets led the comparison on cost and result quality. The evidence remains narrow: Exa designed and ran the benchmarks, each category contains only 20 queries, and no independent evaluation is included. The results support a limited claim about those 40 prompts. Production evaluations should use representative queries and measure valid rows, missed entities, latency, and total cost.
A metered bill replaces the subscription #
New accounts now use consumption-based pricing. According to Exa's pricing reference, charges combine model computation, searches, and optional contact enrichment.
| Component | Price |
|---|---|
| Agent Compute Units | $0.10 per ACU |
| Searches | $5 per 1,000, or $0.005 each |
| Email enrichment | $0.02 per address found |
| Phone enrichment | $0.07 per number found |
An Agent Compute Unit, or ACU, measures the model work used to interpret a query, evaluate candidates, verify matches, and populate enrichment columns. The docs give no public conversion from ACUs to model tokens. In Exa's worked example, 3 ACUs cost $0.30 and 250 searches cost $1.25, producing a $1.55 total. A job returning 50 verified matches therefore costs about $0.031 per match. Websets and Search API charges draw from the same team credit balance.
Limits to budget for
- Variable spend: There is no fixed price per row. Complex criteria, extensive verification, and enrichment can increase ACU usage.
- Irreversible migration: Existing Starter and Pro customers retain their previous pricing until they opt in. After switching to usage-based billing, they cannot return to the old plan.
- Incomplete coverage: Higher recall still leaves gaps. Teams should test coverage before using a Webset as an exhaustive source.
- Data obligations: Citations help with review, while customers remain responsible for licensing, privacy, retention, and outreach requirements in their own applications.
One asynchronous API job #
New users can run searches from the Websets dashboard. Developers can use the exa-py SDK described in the API guide. A request specifies the query, a target result count, and optional enrichment columns for fields that Websets should find for each verified row.
import os
from exa_py import Exa
from exa_py.websets.types import (
CreateEnrichmentParameters,
CreateWebsetParameters,
)
exa = Exa(api_key=os.environ["EXA_API_KEY"])
webset = exa.websets.create(
params=CreateWebsetParameters(
search={
"query": "AI labs focused on large language models",
"count": 5,
},
enrichments=[
CreateEnrichmentParameters(
description="LinkedIn of VP of Engineering",
format="text",
)
],
)
)
webset = exa.websets.wait_until_idle(webset.id)
items = exa.websets.items.list(webset_id=webset.id)
The API runs asynchronously because Websets must find candidates and check each one against the query. The example waits until processing becomes idle before requesting the resulting items. Production code should also inspect the terminal job status and handle timeouts, empty result sets, rate limits, and API errors.
Where verified web sets fit #
Websets suits recurring research tasks that require structured rows, explicit criteria, and source-level evidence. Common applications include:
- Sales and go-to-market: Build account lists from industry, location, product, funding, or hiring signals.
- Recruiting: Find candidates with specific roles, experience, publications, or technical backgrounds.
- Investment research: Track companies, projects, founders, and market activity across fragmented sources.
- Scientific research: Assemble literature sets based on methods, topics, released code, or institutional affiliation.
- Dataset development: Feed sourced examples into training, evaluation, monitoring, or agent pipelines.
At Exa's reported benchmark rates, 300 verified people matches would cost about $5.07, while 300 company matches would cost about $7.20 before additional enrichment and workload-specific variation. Teams replacing a custom search-agent loop can compare both systems on the same queries, using cost per valid row, coverage, latency, and citation quality as the decision criteria.