Exa 发布 3.0 搜索 API，新增 Lite 类型并称 8 项基准领先

Add as a preferredsource on Google

Exa has released Exa 3.0, an upgrade to its web search API for AI agents. The Exa 3.0 search API ships a larger embedding model and a cheaper search type, Exa Lite. Dynamic Highlights is now generally available. Exa claims that it leads Perplexity, Brave and Parallel on 8 benchmarks. Those benchmarks were run by Exa, so you can treat them as self-reported. For developers, the practical change is cost: Lite costs $4 per 1,000 requests.

TL;DR

Size: Exa tracks over 2 trillion URLs and serves 100 billion pages. Embedding model size: Not disclosed.

Runs on: Hosted API only. Exa lists configurable search latency of 450ms to 1s. No self-hosted option.

Performance: Exa 3.0 Auto ranks first on all 8 Exa-run benchmarks against Perplexity, Brave and Parallel.

Best: 91.8% on BrowseComp, plus the widest lead on ATLAS (42.0% vs 37.2% for Perplexity).

Worst: 42.0% on ATLAS is its lowest percentage score. Its BrowseComp lead over Perplexity is only 0.3 points.

What is Exa 3.0?

Exa 3.0 is a search API release from Exa Labs that upgrades retrieval quality across every search type. It is available through the API. Exa runs its own crawler and an independent index. It is not a wrapper on Google or Bing results.

How does Exa 3.0 improve retrieval?

Exa attributes the gains to 3 changes in its retrieval stack:

A larger embedding model: Exa trained it with more compute and data than any earlier model. Parameter count is not disclosed.

Dynamic Highlights, now generally available: The model selects tokens across all results, not page by page. It gives more tokens to better pages and drops repeated information.

Billions of new documents: New coverage includes products, local businesses, flights, events, sports and weather.

According to the Exa’s post, Dynamic Highlights uses 95% fewer tokens than full page text while improving answer quality. Highlights carry no extra charge. That matters for agents, where context tokens often cost more than search calls.

What is Exa Lite?

Exa Lite is a new search type, set with type: "lite" on the /search endpoint, per the changelog. It runs on the same index as Exa Auto. Auto spends more compute on candidate generation, highlights and ranking. Lite supports query, result count, highlights, text, categories, domain and date filters, location and moderation. Unsupported parameters return a 400 error naming the field.

Exa now recommends 3 search types: lite, auto and deep. The fast, instant and deep-reasoning types are legacy but still work.

How does Exa 3.0 perform on benchmarks?

Exa compared each provider’s highest-scoring single-step endpoint: Exa 3.0 Auto, Perplexity Web, Brave LLM Context and Parallel Advanced. All ran in one agentic harness. The agent was GPT-6 Luna, with up to 25 search rounds and 10 results per search. Evaluation happened in October 2026.

The margins vary widely. ATLAS and WebCode show the clearest leads, at 4.8 and 5.6 points over Perplexity. BrowseComp, DSQA and WideSearch leads are each under 1.5 points over the runner-up. ATLAS is a benchmark Exa published itself on October 8, 2026.

On cost, Exa plots agentic search score against cost per task. Lite scores about 79 at about $0.045 per task. Auto scores about 80.5 at about $0.065. Exa says the pair sets a new Pareto frontier among tested providers.

How does Exa 3.0 compare with Perplexity, Brave and Parallel?

FeatureExa 3.0 AutoPerplexity Search APIBrave Search API (LLM Context)Parallel Search API (Advanced)

List price per 1K requests$7 (Lite $4)$5$5$5

BrowseComp (accuracy)91.8%91.5%83.5%63.1%

DSQA (F1)88.2%87.6%78.6%82.8%

WideSearch (F1)56.1%54.0%54.8%53.7%

ATLAS42.0%37.2%28.2%24.8%

WebCode75.5%69.9%68.0%64.5%

FinSearchComp68.0%63.4%57.6%66.0%

Company FindAll (mean entities)15.414.213.39.7

People FindAll (mean entities)7.66.83.83.8

Embedding model sizeNot disclosedNot disclosedNot disclosedNot disclosed

All benchmark scores come from Exa’s Exa 3.0 post and were run by Exa. No vendor has published matching results. Prices come from each vendor’s own pricing page.

What data does the expanded index add?

Exa states that it prioritized personal AI assistant use cases. The official technical page lists 1.27B ecommerce listings with prices and availability, and 211M+ local businesses. It also adds real-time news, sports, weather and market data, along with flights, events, transit and podcasts.

The Exa Library structures this data into searchable collections. These include 320M people pages, 33M company pages and 500M research paper pages. It also lists 19M SEC filings, 10M patents and 9M court opinions.

Where does Exa 3.0 fall short?

Self-run evals: Exa picked the harness, endpoints and agent model. ATLAS is Exa’s own benchmark.

Thin margins: Several leads are under 2 points, which may not survive other harnesses.

Lite limits: Lite rejects unsupported parameters with a 400, so some pipelines need Auto.

Higher Auto list price: Auto costs $7 per 1,000 requests versus $5 for the 3 rivals.

Opaque model details: Embedding model size and architecture are not disclosed.

Key Takeaways

Exa 3.0 leads 8 benchmarks, all run by Exa itself.

Exa Lite costs $4 per 1,000 requests on the same index as Auto.

Dynamic Highlights uses 95% fewer tokens than full pages, at no extra cost.

Biggest leads: WebCode (+5.6) and ATLAS (+4.8) over Perplexity.

Test on your own queries before switching providers.

Michal Sutter+ postsBio ⮌

Michal Sutter is a data science professional with a Master of Science in Data Science from the University of Padova. With a solid foundation in statistical analysis, machine learning, and data engineering, Michal excels at transforming complex datasets into actionable insights.

Michal SutterMeet Drex Max: Nace AI’s Multimodal Decision Model for Computer-Use Agents with a Score 67.59

Michal SutterMeet Grok Voice 3.0: SpaceXAI Voice Agent Model Ranks #1 on Artificial Analysis

Exa 发布 3.0 搜索 API，新增 Lite 类型并称 8 项基准领先

推荐理由

主题

标签

正文