TLDRocket
Sign in

Exa 3.0 Search API Tops Perplexity, Brave and Parallel on 8 Benchmarks

MarkTechPost Michal Sutter ● Covered by 2 sources

Exa says its new search API beats Perplexity, Brave and Parallel on 8 tests. The catch: Exa ran the tests itself, so the brag comes with a built-in asterisk.

Based on reporting by MarkTechPost, Michal Sutter — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Exa has shipped Exa 3.0, a new version of its web search API for AI agents. The company says it improves retrieval across every search type, adds a cheaper option called Exa Lite, and makes Dynamic Highlights generally available. It also says the release leads Perplexity, Brave and Parallel on eight benchmarks, though all of those tests were run by Exa itself.

The practical hook for developers is price. Exa Lite is set at $4 per 1,000 requests, while Exa Auto is listed at $7 per 1,000 requests. Exa says Lite runs on the same index as Auto, but uses less compute. Auto does more work on candidate generation, highlights and ranking, while Lite supports a narrower set of parameters and will return a 400 error if a pipeline sends something it doesn't accept.

Exa’s pitch rests on three changes under the hood. First, a larger embedding model, though the company does not disclose its size or architecture. Second, Dynamic Highlights, which picks tokens across all results instead of page by page and is said to use 95% fewer tokens than full page text while improving answer quality. Third, billions of new documents, including products, local businesses, flights, events, sports and weather.

The benchmark picture is mixed in the way self-run benchmark pictures often are. Exa says Exa 3.0 Auto came first on all eight tests, with the biggest gaps on WebCode and ATLAS. But some margins are tiny. On BrowseComp, Exa’s lead over Perplexity is just 0.3 points. On ATLAS, Exa scores 42.0% versus Perplexity’s 37.2%.

The expanded index is clearly aimed at assistant-style search. Exa says it now covers 1.27 billion ecommerce listings, 211 million-plus local businesses, 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. It also says the system is hosted only, with no self-hosted option, and configurable latency from 450ms to 1s. That's a very particular kind of product: useful, fast, and wrapped in just enough benchmark glitter to make the procurement meeting harder than it should be.

My take — AI-written commentary, not fact-checked reporting

Self-scored benchmarks are the software equivalent of grading your own homework and circling the best bits in red. Exa may well have a strong product here, but the real story is how normal it’s become for AI companies to launch with a leaderboard before anyone else has kicked the tires. Buyers should treat the benchmark chart as a sales slide, not gospel.

Read more about this at: MarkTechPost

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.