Exa has released version 3.0 of its web search API for AI agents, with a larger embedding model, a cheaper new search type and Dynamic Highlights now generally available at no extra charge. According to Exa, the new system beats Perplexity, Brave and Parallel on all eight benchmarks it tested. The result should be read with caution: the tests were run by Exa itself, which chose the agent, endpoints and model used.
Three changes behind the improvement
Exa attributes the gains to three changes in its retrieval pipeline. The first is a larger embedding model, trained with more compute and data than any earlier version; its parameter count has not been disclosed. The second is Dynamic Highlights, now generally available: instead of selecting text page by page, the model picks tokens across all results, giving more space to the better pages and dropping repeated information. The third is an expanded index with billions of new documents covering products, local businesses, flights, events, sports and weather. Exa says Dynamic Highlights uses 95% fewer tokens than full page text, which matters for agents, where context tokens often cost more than search calls.
Lite and Auto
The commercial novelty is Lite, a new search type set with type: "lite" on the /search endpoint, running on the same index as Auto. Auto spends more compute on candidate generation, highlights and ranking; Lite supports query, result count, domain and date filters, location and moderation. Unsupported parameters return a 400 error naming the field. Exa now recommends three search types, lite, auto and deep, while fast, instant and deep-reasoning remain as legacy options.
In the tests, Exa compared each provider's best single-step endpoint: Exa 3.0 Auto, Perplexity Web, Brave LLM Context and Parallel Advanced, all run in one agentic harness with GPT-6 Luna, up to 25 search rounds and 10 results per search. Margins vary widely: ATLAS and WebCode show the biggest leads, 4.8 and 5.6 points over Perplexity; on BrowseComp, DSQA and WideSearch the lead over the runner-up is under 1.5 points. On BrowseComp Exa scores 91.8% against Perplexity's 91.5%, a gap of 0.3 points.
Pricing and index
The list price is $5 per 1,000 requests for Exa Auto's rivals, but Exa 3.0 Auto costs $7 while Lite costs $4. Exa presents the score-versus-cost-per-task trade-off as a new Pareto frontier: Lite reaches a score of about 79 at roughly $0.045 per task, Auto about 80.5 at roughly $0.065. The expanded index also includes personal and everyday data: 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 says it prioritised personal AI assistant use cases, adding real-time news, sports, weather, market data, flights, transit and podcasts.
Exa Neural Search API Returns Structured Contents for LLM Agents Across ...
Limits to consider
The caveats are substantial. Exa picked the harness, endpoints and agent model on its own, and ATLAS is a benchmark created by the company itself. Several leads are under 2 points and may not survive other test setups. Lite rejects unsupported parameters, so some pipelines must stay on Auto. Embedding model size and architecture have not been disclosed. For developers building agents, the practical advice is simple: test Exa on your own queries before switching providers.
No comments yet. Be the first!