Introducing Exa Deep: An Agent for Every Search
Introducing Exa Deep: An Agent for Every Search
Today we're launching a revamped Exa Deep: our best agentic search endpoint. It uses optimized query expansion and LLM reasoning to get high-quality search results. Deep is already used for detailed people & company search, financial research, news reports and more.
This version is faster, cheaper, and adds structured outputs with field-level grounding.
Behind the scenes, Exa Deep uses our recently launched Exa Instant search and LLM reasoning to understand the intent behind your query, generating multiple search agents in parallel, and synthesizing results with citations.
Evals
We benchmarked Exa Deep against Perplexity (Sonar Reasoning Pro) and Parallel Task API (Core, Base) on three challenging evals: HLE-Search, Deep Search QA, and FRAMES.
On every benchmark Exa Deep is up and to the left: faster and more accurate.
How are people using Deep?
These are categories of queries where Deep's capabilities — multi-source synthesis, structured outputs, and grounded citations — are especially valuable:
Financial research. SEC filings, quarterly reports, company intelligence. Queries that require pulling from multiple documents and synthesizing.
Scientific literature. Researchers using it to survey recent publications on a topic. About 20% of Deep queries.
News monitoring. Detailed, ongoing monitoring at set cadences. Deep handles the complexity of tracking evolving stories across sources.
Research agents.If you're building a deep research agent, Deep is the search primitive you want underneath, and Exa Agent runs the whole research loop for you. One API call replaces a complex orchestration layer.
Structured Outputs, Enrichments and Grounding
You can define an output schema, and Deep returns structured JSON with field-level citations:
// Query: Top aerospace companies and their CEO's name
"output": {
"content": {
"companies": [
{
"ceo_name": "Christopher T. Calio",
"company_name": "RTX Corporation"
},
{
"ceo_name": "Kelly Ortberg",
"company_name": "Boeing"
},
// ...
]
},
"grounding": [
{
"field": "companies[0].ceo_name",
"citations": [
{
"url": "https://fintool.com/app/research/companies/RTX/people/christopher-t-calio",
"title": "Christopher Calio - Executive Profile & Compensation"
},
// ...
],
"confidence": "high"
}
]
}Schemas support up to 10 properties in total and 2 levels of nesting. See the output schema limits in the Deep Search guide.
How to use Exa Deep
Set type to deep-lite, deep, or deep-reasoning in your Search API request. Add an outputSchema for structured outputs.
| Tier | Latency | Price |
|---|---|---|
deep | 4-12s | $12/1k requests |
deep-reasoning | 12-50s | $15/1k requests |
Regular deep is 20% cheaper than before.
Try it in the playground, read the Deep Search guide, or see the Exa Deep product page.
Where search is going
Search and agents are converging. The best way to answer a complex query isn't a single search. It's an agent that breaks down the query, searches in parallel, and iterates until it has what it needs.
AIs will soon search the web more than humans, and these AIs need search that can reason, not just retrieve. Exa Deep is our answer.
Since this launch we've shipped Deep Max, our highest-quality agentic search tier, and Exa Agent for long-running research and enrichment.
We're hiring. Come help us build it.
