Exa Cuts AI Agent Search Tokens with Dynamic Highlights

Exa has released a research preview of Dynamic Highlights, a new search primitive designed to optimize how AI agents and retrieval-augmented generation (RAG) pipelines consume web data. Instead of extracting isolated snippets from individual documents, Exa's custom model scans all retrieved pages simultaneously. By concatenating these pages into a single document, the system shifts the token budget toward the most relevant sources, reducing retrieval tokens by an average of 95 percent.

The technology yields significant efficiency gains across different workloads. In single-turn RAG scenarios with a 12,000-character budget, Dynamic Highlights achieved a 40 percent token efficiency gain alongside a 3.8 percent improvement in quality. For multi-turn agentic workflows, the system reduced agent token usage by 30 percent while delivering a 2.1 percent average quality gain. These agentic benchmarks were evaluated using the exa-agent auto model for browsecomp and widesearch tasks, as well as the exa-agent medium model on internal datasets.

Developers can activate the feature by setting the dynamic parameter to true within the highlights block of Exa's Contents API. The feature is integrated into the existing pricing structure, meaning it is free for up to 10 results per search, after which it costs $1 per 1,000 pages.

For AI practitioners, this shift addresses the issue of what Exa calls "context rot," where redundant web data bloats LLM prompts and drives up API costs. By treating token spans, rather than entire documents, as the fundamental unit of retrieval, Dynamic Highlights allows developers to build cost-effective RAG pipelines and deep research agents that require fewer search queries to find precise answers.