Exa 3.0 adds Lite search: check source support before switching

Exa 3.0 gives research teams a new retrieval choice to evaluate: Lite search, alongside Auto and generally available Dynamic Highlights. The October 19 launch changes the configuration behind a source-gathering workflow. It is a reason to review the evidence returned for a task before interpreting a lower request price as a better result.

What changed in the API

Exa’s launch announcement introduces Lite and makes Dynamic Highlights generally available. The company also reports retrieval improvements and broader index coverage. Its benchmark wins and claimed token savings are vendor-reported results; SearchEngineAnswer has not independently reproduced them.

The current pricing table lists Lite at $4 and Auto at $7 per 1,000 base requests with up to ten results. Extra results and AI page summaries have separate charges. Check the selected endpoint and options before projecting the cost of a complete research job.

Compare the answerable task, not just the response size

A shorter excerpt can save processing while omitting the qualification that makes a claim accurate. The useful review unit is therefore a question, the returned sources, and the supported answer. Keep all three together when comparing two configurations.

Consider a fictional product-policy question. Lite returns an excerpt saying returns are accepted; the original page limits that policy to unopened items. If the final answer omits that condition, the reviewer should record an unsupported generalization. A larger excerpt from Auto might resolve it, but the comparison must inspect what each run actually returned.

Change one setting, then inspect disagreements

1. Write a small set of genuine reader questions before collecting results. Include cases with dates, exclusions, or regional conditions.
2. Run the same questions under declared Lite and Auto configurations. Preserve errors and missing results.
3. Open the relevant source pages and compare each excerpt with the nearby qualification.
4. Score factual support separately from URL overlap, latency, and observed cost.
5. Keep the configuration that meets the reader task. Record when a cheaper first pass needs a second retrieval.

This is a proposed evaluation procedure, not an executed benchmark. It does not establish that one search type is universally better. A task involving a simple definition and one involving a disputed policy can justify different retrieval choices.

The Exa Places guide covers the separate problem of interpreting local-business fields. Use the measurement crosswalk when reporting results: retrieved pages, citations, and reader visits still require different evidence.

A study of qualifier-sensitive answer selection examines the next step after evidence arrives: choosing a candidate that satisfies the question’s full requirements. Its experiment assumes the correct candidate is already available.