Exa 3.0 Beats Perplexity on 8 Benchmarks and Slashes Search Costs

Exa, a web search API for applications and AI agents, has released Exa 3.0, which the company describes as its largest search-quality upgrade to date. The release combines a larger embedding model, more efficient result highlights, billions of additional documents, and Exa Lite, a lower-cost tier for high-volume workloads. These changes affect relevance, context-window usage, data coverage, and API costs.

Three upgrades reshape retrieval

Exa attributes the reported gains to three changes that apply across its search modes:

• A larger embedding model: Exa trained it with more data and compute than any of its previous models. Embeddings represent queries and documents as numerical vectors, allowing the system to retrieve pages with related meaning even when their wording differs.
• Dynamic Highlights: The feature is now generally available and selects relevant text across the complete result set. Exa says it uses 95% fewer tokens than sending full-page text while improving answer quality.
• A broader index: Billions of documents now cover products, local businesses, flights, events, sports, weather, and other frequently changing subjects.

Highlights allocate context across results

Dynamic Highlights changes how Exa extracts text for downstream models. Conventional retrieval pipelines often select passages from each page independently, producing repeated facts when several sources cover the same topic. Exa’s model evaluates all returned pages together, allocates more tokens to stronger sources, and removes redundant material.

For retrieval-augmented generation and agent workflows, smaller excerpts can reduce prompt size, model latency, and token charges. Cross-result selection can also fit a broader set of evidence into a fixed context window.

Eight benchmark leads, according to Exa

Exa reports first-place results across eight public and domain-specific evaluations, including coding, finance, company research, and people search. Its published scores place Exa 3.0 ahead of the compared offerings from Perplexity, Brave, and Parallel.

The benchmarks use different tasks and scoring scales, so each row supports comparisons only within its evaluation. Exa published these results; teams considering a switch should also test representative queries from their own workloads.

Lite shares the index and trims compute

Exa Lite and Exa Auto search the same underlying index. Auto spends more compute on candidate generation, which creates the initial pool of possible results, as well as highlight extraction and final ranking. Exa includes the upgraded highlights and expanded data sources in both tiers without an added fee.

The index moves into daily tasks

Exa says its crawler tracks more than 2 trillion URLs and its search infrastructure serves 100 billion pages. The latest expansion grew around customer query patterns, with traffic from personal AI assistants driving demand for commerce, local, travel, and real-time information.

Product, place, and real-time data let one agent search for research papers, check store inventory, find a flight, or retrieve local business hours through the same provider. The expansion moves Exa beyond the long-form web content that established its neural-search product and into workloads often handled by general search or specialized data APIs.

Route requests by the cost of a miss

• Choose Exa Lite for autocomplete, background enrichment, broad discovery, and large batch jobs where request volume dominates the cost calculation.
• Choose Exa Auto for deep-research agents, coding assistants retrieving documentation, financial analysis, and user-facing answers where stronger ranking can justify the higher request price.

Applications with mixed workloads can assign the higher-compute tier to queries where a missing or poorly ranked source carries greater operational cost, while sending routine searches through Lite.

Existing pipelines change automatically

Exa 3.0, Dynamic Highlights, the expanded index, and Exa Lite are available through the API documentation. Existing integrations receive the retrieval and highlight upgrades automatically, so their request schemas require no migration.

Existing integrations may still return different rankings and extracted passages after the rollout. Teams that maintain relevance tests, latency targets, prompt budgets, or cached results should rerun those checks against representative production queries.