Exa 3.0

Exa 3.0 is the largest improvement to search quality that Exa has ever released, delivered as an update to its search API for AI agents. It is built for developers and teams who need to feed AI agents richer, more useful data from the web at a lower cost, and it is available today through the API. The release bundles three core changes: a larger embedding model trained with more compute and data than any previous Exa model, an index expanded with billions of new documents across useful data types, and Dynamic Highlights, which are now generally available. According to the announcement, Exa 3.0 ranks first on every benchmark the team tested, from public evaluations to coding, finance, and go-to-market intelligence. The context behind the release is the rapid growth of AI agents and personal AI assistants that depend on web retrieval to answer questions and complete tasks. Exa states that it has seen an explosion in usage from personal AI assistants, which is why it prioritized expanding its index and data partnerships to serve those use cases. At the same time, AI applications are cost-sensitive: sending full page text to a model consumes large numbers of tokens, and running heavy retrieval for every task adds up. Exa 3.0 addresses both sides of that trade-off at once, pairing broader and more accurate retrieval with a highlight model that dramatically cuts the amount of text passed downstream, plus a new lower-cost search tier for applications where budget is the binding constraint. The result, per the announcement, is a new Pareto frontier for search quality versus cost. The first pillar of the update is the retrieval stack itself. Exa trained a larger embedding model with more compute and data than any previous model, and states that every search type benefits from the improvement. Alongside the model work, the company added billions of documents to its index across data types such as products, local businesses, flights, events, sports, weather, and more. Together these changes produced gains on evaluations spanning coding, finance, deep research, and company intelligence. Because the improvements live in the shared retrieval stack rather than in a single product surface, developers do not need to switch endpoints or rebuild integrations to take advantage of them. The second pillar is Dynamic Highlights, which was previously released as a research preview and is now generally available at no additional charge. Instead of extracting highlights from each page individually, the new model selects tokens across all results, giving more tokens to the best pages, cutting repeated information, and using 95% fewer tokens than full page text. Exa reports that this approach improves answer quality while reducing token consumption, and that improvements to model efficiency have made it substantially faster and cheaper to run. For teams building agents, this means the retrieval layer returns a more compact, ranked selection of relevant text rather than long page dumps, which lowers both cost and context pressure downstream. The third pillar is the expanded index. Exa runs its own crawler and an independent index, tracking over 2 trillion URLs, serving 100 billion pages, and growing around what its customers search for. Exa 3.0 now includes 1.27B ecommerce listings with prices and availability, 211M+ local businesses with hours and ratings, real-time news, sports, weather, and market data, plus flights, local events, transit, and podcasts. This new data adds to coverage customers already rely on, including people, companies, code, and publications. The announcement highlights concrete search examples such as business hours and ratings for places, product prices and availability, game scores, and forecasts, illustrating the kinds of structured, real-world facts agents can now retrieve. Exa 3.0 also introduces Exa Lite, a new search type priced at $4 per 1,000 requests and built for cost-sensitive applications. Lite runs on the same index as Exa Auto. For higher quality, Exa Auto uses more compute across the retrieval stack, including candidate generation, highlights, and ranking, and is priced at $7 per 1,000 requests. Exa positions the two tiers as complementary: Lite offers strong search quality at a lower price, while Auto trades more compute for the highest quality. Measured as agentic search score versus cost per task, Exa 3.0 Lite scores 79 at about $0.045 per task and Auto scores 80.5 at about $0.065, which Exa describes as setting the quality-and-cost frontier among the tested providers. Overall, the product works through a retrieval pipeline that Exa operates end to end: its own crawler feeds an independent index, queries are handled by an embedding model in the retrieval stack, and candidate generation, highlighting, and ranking are applied with differing amounts of compute depending on the search type chosen. Dynamic Highlights then operates across the full result set rather than page by page, deciding how many tokens each result deserves. Developers access all of this through the API, configure which search type they want based on their quality and cost requirements, and get back ranked results with compact highlights and access to the expanded data sources at no additional charge. The stated outcomes center on benchmark performance. Exa 3.0 scores 91.8% on BrowseComp, 88.2% on DSQA, 56.1% on WideSearch, 42.0% on ATLAS, 75.5% on WebCode, 68.0% on FinSearchComp, 15.4 on Company FindAll, and 7.6 on People FindAll, and the announcement states that Exa 3.0 leads Perplexity, Brave, and Parallel on all eight benchmarks. For users, the practical benefits described are higher search quality for agent tasks, richer and more current data from an expanded index, and substantially lower token usage per query, all of which feed into better answers at a lower operating cost. Use cases named in the announcement include coding, finance, deep research, go-to-market intelligence, and company intelligence, which are the areas covered by the benchmark evaluations. The expanded index points to additional scenarios: product search with prices and availability, local business search with hours and ratings, real-time news, sports, weather, and market data lookups, and queries over flights, local events, transit, and podcasts. Exa also explicitly calls out personal AI assistants as a fast-growing usage category driving the index expansion, and cost-sensitive applications as the target for the Lite tier. Availability is immediate: Exa 3.0 is available today through the API, including Exa Lite, upgraded highlights, and the expanded index. Exa Lite costs $4 per 1,000 requests, Exa Auto is $7 per 1,000 requests, and the improved highlights and expanded data sources are included at no additional charge. Developers can get started in the API dashboard or read the documentation, and Exa hosted a live Exa 3.0 Q&A with its Head of Product and Head of R&D on Wednesday, October 21 at 1pm PT. In summary, Exa 3.0 is a search-quality release for AI agents that combines a stronger embedding model, a substantially larger and more structured index, and token-efficient Dynamic Highlights, while adding Exa Lite at $4 per 1,000 requests alongside Exa Auto at $7. Its core value proposition is straightforward: richer data and better search results for agents, delivered at a lower cost per task.