Exa 3.0 adds cheaper agent search, with cached-content limits

A lower-cost search option for AI applications

Exa released Exa 3.0 on October 19, adding a cheaper Lite search mode, expanding its index and announcing general availability for Dynamic Highlights, which selects passages for an AI model to read. The update is available through Exa’s API, giving developers another way to supply web evidence to assistants and agents.

The practical distinction is what happens after a query finds relevant pages. Lite shares Exa Auto’s index, but Exa says Auto spends more computing resources on retrieving, selecting and ranking information. Lite also has fewer controls. The choice affects which evidence an application can retrieve, as well as what it pays to find it.

Source 1Source 3

Lite cannot fetch a fresh copy of a page

Developers select Lite by setting type to lite on the search endpoint. It accepts domain and publication-date filters, location, categories, full text and highlights. But its page content comes exclusively from Exa’s cache. When text or highlights are requested, results without a cached copy are omitted.

Lite rejects requests for generated summaries, structured answer synthesis and live page retrieval. Sending an unsupported parameter produces a 400 error naming the problem; the documented alternative is Auto. A publication-date filter does not establish when a page’s contents were last fetched.

This matters for an assistant checking a changing price or availability statement. Finding a relevant page and obtaining its current contents are separate operations. Our assessment is that Lite suits discovery work where cached evidence is acceptable; applications requiring a fresh fetch need a different request path.

Source 2Source 3Source 8

The base price is only part of the bill

Lite costs $4 per 1,000 requests, compared with $7 for Auto, with up to 10 results included in each request. Additional results cost $1 per 1,000 results. For a hypothetical 100,000 requests returning no more than 10 results each, the base search charges would be $400 for Lite or $700 for Auto, before credits, optional extras or other services.

That calculation covers retrieval, not the language model consuming its output. Separately calling Exa’s Contents endpoint is also a different charge: its pricing documentation lists $1 per 1,000 pages per content type.

Exa offers self-service access with a $10 free-credit balance that resets monthly; completing onboarding earns the first team a one-time $10 bonus. No payment method is required for the free tier. The public plan page lists 10 search queries per second for Starter and up to 25 for Developer.

Source 4Source 9

Highlights reduce context, but rollout documentation disagrees

Dynamic Highlights first appeared as a research preview on August 28. Earlier highlights selected excerpts from each document separately. The newer approach considers the retrieved documents together, allocating more space to useful material and less to repeated information. For an agent reading several accounts of the same event, the intended benefit is less duplication in its input.

Exa reports an average 95% reduction in tokens compared with full page contents. That is a vendor evaluation of retrieved text, not a demonstrated 95% reduction in an application’s total bill.

The October launch says Dynamic Highlights is generally available without an added charge. However, the implementation guide inspected October 22 still calls it a research preview and requires the dated Exa-Beta header. That discrepancy leaves the current request requirement unresolved; we have not tested whether the header can now be omitted.

Source 1Source 5Source 6

More kinds of information in a standard search

Exa’s Library explains how the broader index can serve everyday assistant tasks. Alongside public webpages, Exa uses public and licensed records to assemble structured pages for subjects such as businesses, products, people and markets. Its listed collections include business hours and ratings, retail listings, sports schedules and weather forecasts.

For a hypothetical travel assistant, a venue’s hours and a local weather forecast could be found through the same search integration. That reduces the number of distinct information sources a developer may need to connect. It does not establish that every returned detail is current or that the assistant can make a reservation; those are separate requirements.

Source 7

Search responses now expose weakened restrictions

The October changelog also adds information that can change how an agent interprets its results. Warnings identify ineffective request fields, while notices explain how results were produced. Exa’s guidance tells developers to log warnings and send notices alongside the results to the model.

One notice reports when no matches satisfied a site operator or quoted phrase and Exa removed that restriction. Explicit request-body filters, such as includeDomains, still apply. For an assistant meant to cite an official source, our reading is that this difference is consequential: a successful response alone does not establish that the original query restrictions held.

Existing integrations do not face an announced shutdown in this release. Exa labels fast, instant and deep-reasoning legacy modes but says they continue at their current prices. It says Lite runs the same search as fast, while accepting fewer parameters. The recommended replacements are migration choices; developers moving from fast still need to account for Lite’s smaller parameter set.

Source 2Source 8