Evergreen Tools
|
← Back to AI Tools

Exa 3.0 Search API for AI Agents

A search endpoint built for agents: token-lean highlights, a cheaper tier and far more structured data

Official Source

This page summarises an official announcement — check the vendor site for current availability and pricing.

Features

  • ✓ A larger embedding model trained with more compute and data than any previous Exa model
  • ✓ Dynamic Highlights is generally available, which Exa says uses 95% fewer tokens than full page text
  • ✓ Billions of new documents across products, local businesses, flights, events, sports and weather
  • ✓ New Exa Lite tier at $4 per 1,000 requests, with Exa Auto at $7 per 1,000
  • ✓ First place on eight benchmarks Exa ran itself, including 91.8% on BrowseComp, 88.2% on DSQA and 42.0% on ATLAS

How to Use

  1. Get an API key in the Exa dashboard
  2. Start on the Lite tier by setting type to lite and running your real query set
  3. Compare Auto against Lite on task completion and total cost, since Auto may need only one search where Lite needs two
  4. Switch on Dynamic Highlights to fetch snippets instead of full page text, then measure the input tokens you actually saved

FAQ

What is Exa 3.0?

A new generation of Exa's search API built for AI agents, released in October 2026. It combines a larger embedding model, generally available Dynamic Highlights, a much larger index and a cheaper Lite tier.

How much does it cost?

Exa lists Exa Lite at $4 per 1,000 requests and Exa Auto at $7 per 1,000, with the improved highlights and new data sources included at no extra charge. For comparison, comparable Perplexity, Brave and Parallel endpoints are listed at $5 per 1,000.

Should I trust those first-place benchmark results?

With a discount. All eight benchmarks were run by Exa itself, in an agent harness driven by GPT-6 Luna with up to 25 search rounds and 10 results per search, and ATLAS is a benchmark Exa published itself on 8 October. The margins also vary: ATLAS and WebCode lead Perplexity by 4.8 and 5.6 points, while BrowseComp leads by only 0.3.

Lite or Auto?

Do the arithmetic on completed tasks, not calls. On Exa's own cost-versus-quality chart Lite scores about 79 at roughly $0.045 per task and Auto about 80.5 at $0.065. If Auto resolves a task in one search where Lite needs two, the pricier tier can still come out cheaper.

Does Dynamic Highlights really save 95% of tokens?

That is Exa's figure measured against full page text. Most pipelines already truncate pages, so your real saving will be smaller than that baseline suggests. Measure it on your own query set.