Exa 3.0 wants to make search infrastructure a commodity.
Exa Labs announced the upgraded search API on October 20, 2026, claiming it leads Perplexity, Brave, and Parallel across eight benchmarks. The results are uneven, but the scale behind the service is difficult to ignore: Exa tracks over 2 trillion URLs and serves 100 billion pages.
The benchmark results show a wide spread rather than a clean sweep of excellence. Exa 3.0 scored 91.8% on BrowseComp, 88.2% on DSQA, 56.1% on WideSearch, 42.0% on ATLAS, 75.5% on WebCode, and 68.0% on FinSearchComp.
The remaining tests measure entity discovery: Exa 3.0 returned a mean of 15.4 entities on FindAll and 7.6 mean entities on People FindAll. Its best result was 91.8% on BrowseComp, while its lowest was 42.0% on ATLAS — a useful reminder that “leads on eight benchmarks” does not mean “wins every task by a comfortable margin.”
Exa also says Exa 3.0 Auto ranks first on all eight benchmarks it tested. The comparison figures include 91.5% for Perplexity, 83.5% for Brave, and 63.1% for Parallel, giving Exa a narrow lead over Perplexity in the cited comparison and a wider gap over the other two services.
That gap may matter less than the design of the product. Exa 3.0 uses a larger embedding model, though Exa has not disclosed its size, and its Dynamic Highlights feature uses 95% fewer tokens than full page text. For applications that need useful excerpts instead of entire pages, the token reduction is the more practical feature.
A large index with a hosted-only trade-off
The new index covers far more than ordinary webpages. It includes 1.27 billion ecommerce listings, 211 million local businesses, real-time news, sports, weather, market data, flights, events, transit, and podcasts.
It also includes 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. That range positions Exa 3.0 as search infrastructure for commercial, research, financial, local, and operational use cases rather than a narrow web lookup tool.
The pricing is straightforward. Exa 3.0 Lite costs $4 per 1,000 requests, while Exa 3.0 Auto costs $7 per 1,000 requests.
There is one catch, because there is always one: Exa 3.0 has no self-hosted option and runs only on a hosted API. That keeps deployment simple, but it also leaves control of the search system, index, and service environment with Exa Labs.
Search is becoming the interface layer
The launch arrives as websites add AI chatbots and generative search tools to improve discovery and engagement. Conversational search can reveal customer intent better than traditional menu navigation and keyword search, especially when visitors describe what they want instead of knowing the site’s preferred labels.
That shift changes the role of search from a box on a page into an interface for finding products, businesses, documents, and answers. Exa’s large index and Dynamic Highlights feature target the infrastructure behind that interface, while the pricing gives developers a direct choice between Lite and Auto.
Other model platforms are expanding in parallel. GPTProto added Google’s Gemini 4 Argon to its model catalog on October 16, 2026, four days before Exa 3.0’s announcement.
Gemini 4 Argon provides an API for text reasoning, software engineering, and multi-step workflows. Put together, the releases point to a market assembling around composable AI services: search from one provider, reasoning from another, and website experiences that stitch the pieces into something visitors might actually use.
Exa 3.0 is not a self-hosted search engine, and its benchmark results do not erase the weak ATLAS score. But it combines a 2 trillion-URL crawl, 100 billion served pages, specialized index coverage, lower-token highlights, and a hosted API with clear pricing. That is less glamorous than another chatbot demo — and more useful.
Based on
