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

Search is quietly becoming the most important plumbing in agentic AI. Every agent that answers questions about the live world needs to fetch pages, and the quality and cost of that retrieval shape both accuracy and the bill. Exa has shipped Exa 3.0, with a larger embedding model, generally available Dynamic Highlights, billions of new documents, and a cheaper search type called Exa Lite at $4 per 1,000 requests.

Exa claims its Auto mode ranks first on all eight benchmarks it ran against Perplexity, Brave and Parallel, including 91.8% on BrowseComp. The practical story for builders may be less the leaderboard and more the cost: Dynamic Highlights reportedly uses 95% fewer tokens than full page text at no extra charge, and for agents, context tokens often cost more than the search call itself.

We want independent search indexes to thrive, and Exa running its own crawler over more than 2 trillion tracked URLs, rather than wrapping Google or Bing, is good for the ecosystem. A web where every AI agent sees the internet through one or two gatekeepers would be a narrower, more fragile web. So we root for credible challengers here.

That is exactly why the benchmark framing deserves scrutiny. Exa chose the harness, the endpoints and the agent model, and ATLAS — where it shows one of its widest leads — is a benchmark Exa itself published earlier this month. On BrowseComp, the lead over Perplexity is 0.3 points; on DSQA and WideSearch the margins over the runner-up are each under 1.5 points. Those are differences that could flip under a different harness or query set. The cleaner signals are WebCode and ATLAS, at roughly five-plus points, and even there no rival has published matching results.

The pricing picture is more nuanced than 'cheaper.' Lite is $4 per 1,000 requests, but Auto lists at $7, above the $5 charged by all three rivals. Exa's argument is that cost per task, not per request, is what matters, and its chart puts Lite around $0.045 and Auto around $0.065 per task. That is a reasonable way to think about agent economics, and token-efficient highlights are a real advantage — but it is still Exa's own math.

The index expansion is where we'd linger. Exa says it prioritized personal assistant use cases and structured collections including hundreds of millions of people pages and company pages, along with court opinions, patents and SEC filings. Better retrieval for flights, local businesses and products is useful. Hundreds of millions of indexed people pages, served to autonomous agents, is also an obvious privacy and profiling surface. We would like to see Exa explain what controls exist on people search and how individuals can correct or remove records.

Exa now recommends three search types — lite, auto and deep — with older types still functioning as legacy. Lite runs on the same index as Auto but spends less compute on candidate generation and ranking, and rejects unsupported parameters with a 400 error. Exa's evaluations used GPT-6 Luna as the agent with up to 25 search rounds in October 2026.