Exa Labs has released the Exa 3.0 search API, which utilizes its own crawler and index rather than relying on Google or Bing. The update introduces a new search type called Exa Lite, designed to optimize token usage. According to the company, the Dynamic Highlights feature consumes 95% fewer tokens than full-page text while maintaining answer quality at no additional cost.
The company now categorizes its search offerings into three types: lite, auto, and deep. The lite version supports various filters, including domain, date, location, and moderation. Unsupported parameters in the API will trigger a 400 error. The previous fast, instant, and deep-reasoning search types remain functional as legacy options.
In October 2026, Exa conducted internal benchmarks comparing Exa 3.0 Auto against Perplexity Web, Brave LLM Context, and Parallel Advanced using the GPT-6 Luna agent. Exa reports that Exa 3.0 outperformed these competitors across eight benchmarks, including ATLAS and WebCode, where it led by 4.8 and 5.6 points respectively.
Exa analyzed the cost-efficiency of its services, noting that the Lite model scores approximately 79 at a cost of $0.045 per task, while the Auto model scores 80.5 at $0.065 per task. The company claims these figures establish a new Pareto frontier in search performance relative to cost.
The Exa Library currently provides access to structured data collections, including 320 million people pages, 33 million company pages, and 500 million research papers. Additionally, the index contains 19 million SEC filings, 10 million patents, and 9 million court opinions.
