AI search API company Exa — which recently launched its Deep Max search API — just shipped a significant upgrade to its Highlights feature, the tool that extracts query-relevant dense summaries from web pages. According to the company, a 500-character Highlights snippet now matches the accuracy of an 8,000-character full page, all while using 16x fewer tokens.
The summaries are generated in real-time per request (no caching) and complete in under 100 milliseconds. On the SimpleQA benchmark, 500-character Highlights performed as well as 8,000-character full-text, and 4K-character Highlights beat 32K-character full-text. The gains are most dramatic in long-document contexts like technical code docs, API references, SDK documentation, and research papers: with a 500-character budget, Highlights hit 60% accuracy versus just 6% for the full text.
Exa says its own agent endpoints — /answer, Deep, and Websets — already use Highlights under the hood instead of reading raw page content. The upgrade is purely a model-level improvement; any existing API call with the enableHighlights parameter will automatically get better results without code changes.