Exa has released the Exa 3.0 search API, introducing a larger embedding model, a cheaper Exa Lite search tier, and generally available Dynamic Highlights to improve web retrieval for AI agents. According to company-run benchmarks tested in October 2026, Exa 3.0 Auto outperforms Perplexity, Brave, and Parallel across eight different evaluation tasks, achieving a top score of 91.8% on BrowseComp.
The upgraded retrieval stack features a new larger embedding model trained with increased compute and expanded data coverage, including over 1.2 billion e-commerce listings and hundreds of millions of people, company, and local business pages. Additionally, the new Dynamic Highlights feature intelligently selects tokens across all results to reduce token usage by 95% compared to full page text. Developers can access the new capabilities via the hosted API, with the cost-effective Exa Lite tier priced at $4 per 1,000 requests.
This release matters because reducing token consumption and improving search retrieval accuracy are critical factors for cutting operational costs and enhancing performance in autonomous AI agents.