Loading page…
Lite and Auto share an index but use different amounts of compute. Exa also says its upgraded highlights use 95% fewer tokens than full-page text.
Listen to this story
Exa’s October 19, 2026, 3.0 release adds Exa Lite at $4 per 1,000 requests alongside $7 Exa Auto, with both tiers querying the same index; Auto uses more retrieval compute rather than a broader corpus. The update also makes Dynamic Highlights generally available and adds shopping, local-business, and real-time data coverage, while Exa says highlights use 95% fewer tokens than full-page text. Its benchmark lead is company-reported and limited to eight evaluations, so buyers should test Lite-versus-Auto quality on their own workloads.
On Exa’s agentic-search chart, Lite scores about 79 at $0.045 per task versus Auto’s 80.5 at $0.065; those are task costs, not API rates.
New coverage includes 1.27 billion ecommerce listings, more than 211 million local-business listings, and real-time news, sports, weather, and market data.
Exa reports scores of 91.8% on BrowseComp, 88.2% on DSQA, 56.1% on WideSearch, and 42.0% on ATLAS; its first-place claim covers only the eight benchmarks it tested.
Exa released Exa 3.0 on October 19, 2026, adding a $4-per-1,000-request search option for applications and AI assistants. Available immediately through its API, the release combines a larger search index with more selective page excerpts. Exa also claims first place on eight benchmarks, putting its quality claims alongside a concrete lower-priced option.
The update has three main parts: a larger model for finding relevant material, a new way to select text from search results, and billions of additional documents. Exa says every search type benefits from those retrieval improvements. The upgraded highlights and expanded data sources carry no additional charge.
The model upgrade sits underneath those search options. Exa says it trained a larger embedding model—the model used to find relevant material—with more data and computing power than any of its predecessors. It attributes the retrieval gains to that work together with the highlights and index changes.
Exa Lite is the new lower-cost search type. It searches the same underlying index as Exa Auto, so the distinction is not access to a separate, smaller collection of pages. Instead, Auto spends more compute on finding possible results, selecting highlights and ranking the material returned.
The new cost-sensitive option uses the same index as Auto, with less compute across the retrieval process.
Exa positions Auto as the higher-quality option, using more compute for candidate generation, highlights and ranking.
That makes Lite a new price choice, not an announced price cut to Auto. Exa presents the two as alternatives for different budgets: Lite for cost-sensitive applications, Auto for higher quality. Both include the expanded data sources and improved highlights; the additional retrieval compute is what separates their approaches.
Exa’s agentic-search chart illustrates its claimed tradeoff: Lite scores about 79 at roughly $0.045 per task, while Auto scores about 80.5 at roughly $0.065. Those plotted costs are per task, not the published per-request API rates. In that comparison, Auto’s additional spending buys a smaller increase in the measured score; it does not establish the difference for every application.
Dynamic Highlights, previously a research preview, is now generally available. Rather than extracting highlights from each page separately, the new model selects tokens—small units of text—across the full set of results. It allocates more text to the best pages and cuts repeated information.
Exa says this uses 95% fewer tokens than full-page text while improving answer quality. The comparison is specifically against sending complete page text, not against the previous highlights feature. Exa also says model-efficiency improvements have made Dynamic Highlights substantially faster and cheaper to run.
The index expansion addresses a different part of the search process: what information is available to retrieve. Exa says rising use by personal AI assistants shaped its priorities for new coverage and data partnerships. The additions reach beyond its existing coverage of people, companies, code and publications.
Exa runs its own crawler and an independent index. It says that infrastructure tracks more than 2 trillion URLs and serves 100 billion pages. Those are separate measures of its overall search infrastructure, rather than counts of the documents added in this release.
Exa’s published benchmark chart places 3.0 ahead of the compared Perplexity, Brave and Parallel offerings on all eight evaluations. The tests cover several kinds of work, including coding, finance, deep research and company intelligence. The first-place assertion applies to the benchmarks Exa tested, not every possible search task.
Among the company-published results are 91.8% on BrowseComp, 88.2% on DSQA, 56.1% on WideSearch and 42.0% on ATLAS. Its chart also reports results for WebCode, FinSearchComp, Company FindAll and People FindAll. Those scores provide specific comparisons behind the launch’s headline claim, while remaining Exa’s account of its performance.
Loading discussion...
Join the conversation
Explain when less source material would help—or leave you uneasy.
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.