Exa 3.0 Search API Claims Superiority Over Perplexity, Brave on Benchmarks

Exa Labs has released Exa 3.0, a significant upgrade to its web search API designed specifically for AI agents. The release introduces a new lower-cost search tier called Exa Lite and makes the Dynamic Highlights feature generally available for all users.

According to MarkTechPost, Exa claims its 3.0 Auto search type outperforms competitors Perplexity, Brave, and Parallel on eight distinct benchmarks. These benchmarks were conducted and reported by Exa itself, meaning the performance figures should be treated as self-reported data rather than independent verification.

Technical Improvements and Performance

Exa attributes the reported performance gains to three primary changes in its retrieval stack. First, the company deployed a larger embedding model trained with more compute and data than previous versions, though the specific parameter count remains undisclosed. Second, Dynamic Highlights is now generally available, allowing the model to select tokens across all results rather than processing pages individually.

This highlighting mechanism allocates more tokens to higher-quality pages while dropping repeated information. Exa states that Dynamic Highlights uses 95% fewer tokens than full page text while improving answer quality. These highlights carry no extra charge, which is significant for AI agents where context tokens often cost more than the search calls themselves.

The third factor is the addition of billions of new documents to the index. This expanded coverage includes products, local businesses, flights, events, sports, and weather data. Exa operates its own crawler and an independent index, meaning it is not a wrapper around results from Google or Bing.

Benchmark Results and Pricing

On the eight benchmarks run by Exa, the 3.0 Auto search type ranked first in every category. The highest score was 91.8% on BrowseComp, where it led Perplexity by a narrow margin of 0.3 points. The widest lead occurred on the ATLAS benchmark, where Exa scored 42.0% compared to 37.2% for Perplexity.

The 42.0% score on ATLAS represents the lowest percentage achieved by Exa 3.0 across the tested metrics. For developers, the most immediate practical change is the introduction of Exa Lite. This new search type costs $4 per 1,000 requests, set via the type: "lite" parameter on the /search endpoint.

Exa Lite runs on the same index as Exa Auto but consumes less compute for candidate generation, highlights, and ranking. It supports standard filters including query, result count, highlights, text, categories, domain, and date. Location and moderation filters are also supported, while unsupported parameters return a 400 error naming the specific field.

Exa now recommends three primary search types: lite, auto, and deep. The previously available fast, instant, and deep-reasoning types are considered legacy but remain functional. The API is hosted only, with configurable search latency ranging from 450ms to 1s. Exa tracks over 2 trillion URLs and serves 100 billion pages.

What should you do?

Developers integrating AI agents should evaluate the cost-performance tradeoff of the new Exa Lite tier against existing solutions. Since the benchmarks are self-reported, independent testing is advisable before committing to large-scale integration. Monitor token usage closely, as Dynamic Highlights may reduce context costs significantly for certain query types.

What to do and how to stay safe: Exa

• Test the Exa Lite API endpoint with your specific query types to verify the $4 per 1,000 requests pricing and performance stability.
• Compare the self-reported benchmark results with your own internal testing against current search API providers like Perplexity or Brave.
• Implement Dynamic Highlights in your agent workflows to measure the claimed 95% reduction in token usage for context generation.
• Review your current search latency requirements, as Exa offers configurable latency between 450ms and 1s for its hosted API.

Step-by-step guide: Mobile Device Security: How to Protect Data Beyond Passwords

General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.

Frequently asked questions

How much does the new Exa Lite search type cost?

Exa Lite costs $4 per 1,000 requests, according to the release details reported by MarkTechPost.

Are the performance benchmarks independent?

No, the benchmarks were run by Exa itself, so the results leading Perplexity, Brave, and Parallel are self-reported.

Is Exa 3.0 a wrapper for Google or Bing?

No, Exa runs its own crawler and maintains an independent index, rather than wrapping results from other search engines.