Less text returned, more data indexed: Exa 3.0 claims top spot across eight benchmarks

Leading eight internally tested benchmarks, Exa 3.0 adds a $4 Lite offering and snippets 95% shorter than full pages to its agent engine.

Provider of a web search API designed for AI agents, Exa relies on its own crawler, which tracks over 2 trillion URLs and serves 100 billion pages according to the company. Version 3.0 combines a larger embedding model, a new method for extracting useful passages, and an index expanded with practical data. From this, the startup draws a strong claim: first place on each of the eight benchmarks it tested, competing against search APIs from Perplexity, Brave, and Parallel. ## Eight first places, all measured internally The launch post displays 91.8% on BrowseComp, 88.2% on DSQA, 75.5% on WebCode, 68% on FinSearchComp, 56.1% on WideSearch, and 42% on ATLAS, along with two exhaustive search tests for people and companies. The scores of the three competitors, which the chart caption states are all lower, are not detailed in the text, nor are the protocol or the configuration selected for each. As noted by Startup Fortune, no raw output or reproducible test bench has been published. One of the eight tests, ATLAS, was designed by Exa itself. Its 547 tasks require finding all entities that match given criteria and then completing their attributes, with a row in the final table counted as correct only if every one of its cells is. To explain its gains, the company highlights its larger embedding model, trained with more compute and data than previous ones, without sharing its size or architecture. ## Highlights chosen token by token across all results The most concrete change involves the excerpts, or highlights, that the API returns instead of full pages. Rather than parsing each page separately, a dedicated model assembles all retrieved documents into a single text and selects the relevant tokens from it in a single pass. This global reading allows it to allocate more space to the best sources and discard repeated information from one result to another. The claimed gain reaches an average of 95% fewer tokens compared to the full text of the pages, though the launch post does not detail the measurement. The post dedicated to these dynamic highlights, published during their research preview, provides more modest figures at the agent level. On BrowseComp, WideSearch, and internal datasets, Exa's agent consumed about 30% fewer tokens for an average quality increase of 2.1%, with highly variable savings depending on the tasks. Compared to traditional highlights in single-query search with a fixed budget of 12,000 characters, the company measures a 40% efficiency gain per token and a 3.8% higher quality. Version 3.0 now presents these highlights as included at no extra charge. ## A $4 Lite plan, served from cache The new Lite search type costs $4 per 1,000 requests for up to ten results, plus $1 per 1,000 results beyond that, compared to $7 for Auto mode. Both query the same index, but Lite dedicates less compute to each retrieval step, and the changelog specifies that it only serves cached content, without ever fetching a live page. A request sending it an unsupported parameter returns a 400 error, a point to verify before any migration. The former Instant type, at the same price, moves to legacy status without being removed. On an agentic search chart whose benchmark is unnamed, the company places Lite at 79 points for roughly $0.045 per task and Auto at 80.5 points for roughly $0.065. According to the company's description, this extra cost corresponds to the additional compute Auto mode devotes to candidate generation, highlights, and ranking. ## 1.27 billion product listings in the index The index expands to include data designed for AI assistants: 1.27 billion product listings with pricing and availability, over 211 million businesses with hours and ratings, flights, local events, transit, and podcasts, as well as real-time data on news, sports, weather, and markets. The collections library also lists 500 million scientific papers, 320 million people profiles, and 33 million companies, all accessible through a standard search call. These additions are included in the query pricing, and local business results surface automatically in Auto and Fast modes.