According to Watchful AI monitoring, Exa has released the Exa Agent Web Research AI API, which executes deep web research and data enrichment tasks at a cost of over 50% less than cutting-edge models such as GPT 5.5 and Opus.
The Exa Agent achieves high token efficiency retrieval by combining a large language model with the Exa web search engine and the Highlights text extraction model. When dealing with large-scale datasets, the system divides the task into multiple subtasks, assigns multiple sub-agents for concurrent retrieval of different domains. During the research process, the system dynamically integrates state-of-the-art models with low-cost models and uses the Highlights mechanism to extract core webpage snippets, achieving a maximum reduction of 94% in token consumption, significantly reducing inference latency and costs.
In the WideSearch benchmark test measuring webpage data extraction and structuring, the Exa Agent excels in balancing row-level F1 score and query cost. Compared to competitors such as Perplexity Agent and Parallel Task, the Exa Agent can drastically reduce the cost per query while maintaining high accuracy. The development team points out that the traditional cell-level F1 evaluation tends to be tolerant of data misalignment errors, hence WideSearch adopts a stricter row-level F1 evaluation, considering a query successful only when all attributes in the entire row match the correct entity.
The Exa Agent has now opened its API service and offers five levels of computational intensity modes. The basic minimal mode charges $0.012 per request, the default medium mode charges $0.10, and the highest-intensity xhigh mode charges $1.00. Users can also enable the auto mode for the system to dynamically allocate computing power. Furthermore, the API interface supports custom JSON templates for direct structured table output, allowing developers to import their own data lists for the agent to autonomously retrieve and enrich background information from the web.
