Image: Exa

Exa has released Exa 3.0, which it calls the biggest jump in search quality it has ever shipped. Exa is a search engine built for AI apps and agents. Developers plug it into chatbots, coding tools and research assistants so the AI can look things up on the web.

Exa announced the update on X and in a blog post. The company says Exa 3.0 is now first on every benchmark it tested. Those tests cover coding, finance, deep research and company data. Exa ran these tests itself, and no outside group has checked the results yet.

Introducing Exa 3.0: our largest ever improvement to search quality.

Exa is now #1 on every benchmark we tested, from public evals to verticals like coding and GTM intelligence.

It's the result of multiple research breakthroughs 🧵 pic.twitter.com/3qxYZ9nYKz

— Exa (@ExaAILabs) October 19, 2026

Most people will never visit Exa directly. They meet it through apps that use it in the background. When an AI assistant fetches a product price or a news story, a search API like Exa often does the fetching. A better search layer can mean fewer wrong answers in the apps you already use.

Exa says the gains come from three changes. It trained its largest embedding model yet. An embedding model helps a search engine match a question to pages by meaning instead of exact words. Exa also made a feature called Dynamic Highlights generally available at no extra charge.

Highlights matter for cost. An AI model pays for every word it reads, counted in tokens. Exa says its new highlights use 95% fewer tokens than full page text. The model picks text across all results, gives more space to the best pages and drops repeated information. Exa says this also improves answer quality, which is its own claim.

The third change is more data. Exa says it runs its own crawler and tracks over 2 trillion URLs. It has added 1.27 billion shopping listings with prices and stock status. It also added more than 211 million local businesses with hours and ratings. Live news, sports, weather and market data are in the index too, plus flights, local events, transit and podcasts. Exa says personal AI assistants drove this push.

The new Exa Lite costs $4 per 1,000 searches. The higher-quality Exa Auto costs $7 per 1,000. Both run on the same index, and Auto spends more computing power on each query. Exa says all of it is live in its API today.

Lite is not quite a price cut. Exa's pricing page lists its older Instant search at the same $4 rate. The change is that Lite runs on the improved index and model. Rivals sit close by. Perplexity's Search API costs $5 per 1,000 requests, or $1 with its Fast Search option. Tavily charges $0.008 per credit on pay as you go, which works out to $8 per 1,000 basic searches.

Search is now a busy market for AI companies. We recently covered how Firecrawl, a web scraping tool for AI, landed on Google Cloud Marketplace.

Practical tips for developers

Try Lite first if you run lots of simple lookups, since it is the cheapest Exa option. Exa gives new accounts $10 in free credits, and it says the free balance resets each month. That covers up to 2,500 Lite searches.

Run your own test before switching. Pick 20 to 50 real questions from your app and compare Exa, Perplexity and Tavily side by side. Count wrong answers and total cost, since vendor benchmarks may not match your use.

Turn on highlights if you pay for model tokens. Exa says they cost nothing extra, so they can cut your model bill.

Exa will hold a live Q&A about Exa 3.0 on October 21 at 1pm Pacific time, which is October 22 at 1:30am IST.