Exa Releases 3.0 Search API, Claims Benchmark Lead Over Perplexity, Brave and Parallel

Exa 3.0 Auto ranked first on all eight Exa-run tests, the company said. On BrowseComp, it scored 91.8% against 91.5% for Perplexity, 83.5% for Brave and 63.1% for Parallel, a 0.3-point lead over Perplexity. On ATLAS, a benchmark Exa published on October 8, 2026, Exa scored 42.0%, compared with 37.2% for Perplexity, 28.2% for Brave and 24.8% for Parallel. Exa also reported 88.2% on DSQA versus 87.6% for Perplexity, 56.1% on WideSearch versus 54.0%, 75.5% on WebCode versus 69.9%, and 68.0% on FinSearchComp versus 63.4%. In entity-finding tests, Exa averaged 15.4 companies against 14.2 for Perplexity and 7.6 people against 6.8.

MarkTechPost reported that the margins vary. The clearest leads were 4.8 points on ATLAS and 5.6 points on WebCode over Perplexity; the leads on BrowseComp, DSQA and WideSearch were each under 1.5 points. Exa compared each provider's highest-scoring single-step endpoint, Exa 3.0 Auto, Perplexity Web, Brave LLM Context and Parallel Advanced, in one agentic harness using GPT-6 Luna, with up to 25 search rounds and 10 results per search. The evaluation took place in October 2026.

The release adds a larger embedding model, whose parameter count Exa did not disclose, and makes Dynamic Highlights generally available. Exa said Dynamic Highlights selects tokens across all results rather than page by page, giving more tokens to better pages and dropping repeated information. The company said it uses 95% fewer tokens than full page text while improving answer quality and carries no extra charge; Exa said that matters for agents because context tokens often cost more than search calls. Exa Lite, a new search type set with type lite on the /search endpoint, runs on the same index as Exa Auto but spends less compute on candidate generation, highlights and ranking. Lite supports query, result count, highlights, text, categories, domain and date filters, location and moderation; unsupported parameters return a 400 error naming the field.

Pricing is central to the release. Exa Lite costs $4 per 1,000 requests, while Exa 3.0 Auto costs $7 per 1,000 requests, compared with $5 per 1,000 for Perplexity, Brave and Parallel, according to each vendor's pricing page. Exa plotted agentic search score against cost per task: Lite scored about 79 at about $0.045 per task, and Auto scored about 80.5 at about $0.065. Exa said the pair sets a new Pareto frontier among the tested providers.

Exa said its index tracks more than 2 trillion URLs and serves 100 billion pages. The expanded coverage includes 1.27 billion ecommerce listings with prices and availability, 211 million local businesses, real-time news, sports, weather and market data, as well as flights, events, transit and podcasts. The Exa Library also lists 320 million people pages, 33 million company pages, 500 million research paper pages, 19 million SEC filings, 10 million patents and 9 million court opinions. Exa lists configurable search latency of 450 milliseconds to 1 second and offers the service only as a hosted API, with no self-hosted option.

Exa now recommends three search types: lite, auto and deep. The fast, instant and deep-reasoning types are legacy but still work, the company said. Exa runs its own crawler and an independent index rather than wrapping Google or Bing results.

MarkTechPost noted several limits. Exa chose the evaluation harness, endpoints and agent model, and ATLAS is Exa's own benchmark. Several leads are under 2 points and may not survive other harnesses. Lite rejects unsupported parameters with a 400 error, so some pipelines need Auto, whose list price is $7 per 1,000 requests against $5 for the three rivals. Exa did not disclose the embedding model's size.