Evidence at a glance
The Upgrade Makes Retrieval Spend Configurable
Exa Labs released Exa 3.0 on October 19, 2026. It is an upgrade to the company’s web search API for AI agents, adding the lower-priced Exa Lite search type and making Dynamic Highlights generally available. Exa describes the release as an improvement to retrieval quality, but the more concrete engineering change is that retrieval effort now has tiers. Lite and Auto use the same index, while Auto spends more compute on candidate generation, highlight extraction, and ranking.
That means choosing a search service does not have to be only a question of which provider ranks highest. For an agent that searches repeatedly and then passes results to a model, retrieval cost, context usage, and result quality all affect the end-to-end task. Exa 3.0 tries to make part of that trade-off adjustable: developers can use Lite to control the price of a request, or choose Auto when they want more processing in the retrieval path. This is a budget-tiering proposal, not a guarantee that either mode will suit a particular workload.
A Shared Index, with Compute Buying a Marginal Gain
Exa’s published cost-versus-score chart gives a concrete example of the tiering. Lite costs about $0.045 per task and scores roughly 79, while Auto costs about $0.065 per task and scores about 80.5. On the chart’s terms, the extra $0.02 or so buys an increase of around 1.5 points. The higher tier therefore does not improve quality for free. It spends more on retrieval processing for a limited additional gain.
Those per-task costs should not be confused with the API’s price per thousand requests. Exa lists Lite at $4 per 1,000 requests and Auto at $7. The report lists Perplexity, Brave, and Parallel at $5 per 1,000 requests. A task can involve several searches, and teams will differ in how they use results, so request-level price and end-to-end agent-task cost need to be tracked separately. A cheap individual call can otherwise obscure the total cost of repeated searches or downstream context handling.
Highlights Shift Search from Finding Pages to Organizing Evidence
Dynamic Highlights changes how results enter an agent’s context. Exa says the system does not extract a separate passage from each page. Instead, it selects content across the full set of results, gives more tokens to more useful pages, and removes repeated information. The company reports that Highlights uses 95% fewer tokens than full-page text while improving answer quality, with no additional charge. These are vendor-reported results, not a basis for assuming the same savings or quality gain for every query.
The mechanism targets a practical problem: finding a web page does not mean an agent has received usable evidence. Full pages consume context, while page-by-page extraction can send similar information to the model more than once. Allocating summary space across results asks the retrieval system to do some filtering and deduplication itself. Teams building agents should measure more than answer correctness. They should also check how much text is passed along, whether the evidence covers distinct sources, and whether context savings are offset by additional search rounds.
Eight Benchmark Leads, but Margins Vary
Exa says Auto ranks first on all eight benchmarks, but Exa ran the tests itself. The comparison covered Exa 3.0 Auto, Perplexity Web, Brave LLM Context, and Parallel Advanced. All providers ran in the same GPT-6 Luna agent harness, with up to 25 search rounds and 10 results per search, and the evaluation took place in October 2026. A shared harness makes comparison within that setup more useful, but it is not an independent replication. Nor does it show that the results would hold with a different agent or query set.
The margins matter as much as the headline. On BrowseComp, Exa scores 91.8% against Perplexity’s 91.5%, a gap of only 0.3 percentage points. DSQA is 88.2% versus 87.6%, and WideSearch is 56.1% versus Brave’s 54.8%. Larger gaps appear on ATLAS, where Exa scores 42.0% against Perplexity’s 37.2%, and WebCode, where it scores 75.5% against 69.9%. On FinSearchComp, Exa scores 68.0% and Parallel 66.0%. Company FindAll reports an average of 15.4 entities for Exa versus 14.2 for Perplexity, while People FindAll reports 7.6 versus 6.8. ATLAS was published by Exa, making independent evaluation particularly important before treating it as evidence of broader retrieval capability.
Index Expansion and Hosting Limits Shape the Fit
Exa says version 3.0 uses its own crawler and independent index rather than wrapping Google or Bing results. The company says the index tracks more than two trillion URLs and serves 100 billion pages, with new coverage for products, local businesses, flights, events, sports, and weather. Its published details include about 1.27 billion product records and more than 211 million local businesses. Those figures suggest that Exa is expanding search into a data entry point for personal assistants. Scale alone, however, does not establish the freshness or accuracy of individual records, or reveal what the index misses.
There are clear limits on deployment and integration. Exa is available only as a hosted API, with no self-hosted option, and the embedding model’s parameter count and architecture have not been disclosed. Lite and Auto share an index, but Lite returns a 400 error identifying any unsupported parameter. Teams moving from another search type therefore need to check request parameters and failure handling. A prudent evaluation would compare Lite, Auto, and the existing service on the team’s own queries and agent workflow, measuring accuracy, latency, cost per task, and data freshness. That evidence can show which queries justify Auto’s heavier processing. If a workload or governance requirement demands more control than a hosted API provides, the current deployment model may simply be a poor fit.