Exa 3.0 search token consumption reduced by 47%, independent evaluation matches Perplexity.

Beating AI News Flash: AI search company Exa has released Exa 3.0, primarily providing web search capabilities for AI Agents. Exa claims that the new version ranks first across all 8 evaluations, including deep research, code retrieval, finance, and enterprise investigation, surpassing Perplexity, Brave, and Parallel, which participated in the tests.

Third-party organization Artificial Analysis tested using the same model and Agent framework and found that Exa Auto achieved a comprehensive score of 80, tying for first place with Perplexity Search. However, Exa takes an average of 43.7 seconds to complete a single task, while Perplexity requires only 25.5 seconds. The lower-priced Exa Lite scored 79, with time shortened to 29 seconds.

The focus of this upgrade is to reduce the cost of AI reading search results. The new Dynamic Highlights no longer extracts content page by page but comprehensively analyzes all search results, allocating more Tokens to the most valuable snippets while removing duplicate information. Exa claims that compared to reading full web pages, this can reduce Token consumption by 95%. Artificial Analysis's actual testing shows that compared to the old summary mode, the total Tokens per task for the Agent dropped from 299,000 to 157,000, a reduction of 47%, and total cost decreased by 21%. This technology was already available in preview in August and is now officially launched.

Exa has also trained a larger Embedding model and expanded its search index, currently covering 1.27 billion e-commerce product listings, over 211 million merchants, and real-time data such as flights, weather, and sports events. The newly launched Lite tier charges $4 per thousand searches, and Auto charges $7, both now available via API.