Exa Launches Revamped Agentic Search API

Exa is launching a revamped version of its agentic search product, Exa Deep. The new endpoint uses LLM reasoning and parallel query expansion to return structured, cited answers from across the web, and the company says it's faster and cheaper than the previous version.

Exa, which came through YC's Summer 2021 batch with early seed backing from Pioneer Fund, has since raised an $85M Series B led by Benchmark at a $700M valuation, with participation from Lightspeed and Nvidia's NVentures.

Exa Deep works by breaking a complex query into multiple sub-queries, running them in parallel through Exa's search infrastructure, and then synthesizing results with citations. Under the hood, it builds on Exa Instant, the company's recently launched low-latency search endpoint, layered with LLM reasoning to understand query intent.

The key addition in this version is structured outputs with field-level grounding. Developers can define an output schema and get back structured JSON where each field comes with its own citation and confidence score. That's useful for applications like financial research, where you need to know exactly where a data point came from, or scientific literature surveys, where traceability matters.

Exa says roughly 20% of Deep queries today come from researchers surveying recent publications. Other common use cases include SEC filings and company intelligence, news monitoring at set cadences, and as the search primitive underneath other deep research agents.

The company benchmarked Exa Deep against Perplexity's Sonar Reasoning Pro and Parallel Task API, claiming it's both faster and more accurate. Pricing starts at $12 per 1,000 requests for the standard tier and $15 per 1,000 for the reasoning tier, with the standard tier now 20% cheaper than before.

The thesis behind Exa Deep is straightforward: as AI agents do more of the web searching that humans used to do, they need search infrastructure that can reason, not just retrieve. A single API call to Exa Deep is meant to replace what would otherwise require a complex orchestration layer of multiple search calls and post-processing.

The new endpoint is available now via Exa's API dashboard.