Exa announced Exa 3.0 on Oct. 19, saying the update was available through its Search API. The company also made Dynamic Highlights generally available at no additional charge and introduced Exa Lite, a lower-priced search type.
Dynamic Highlights determines how much text to return from each search result within a shared budget. Rather than extracting highlights from each page separately, Exa says its model selects text across all results, allocating more to the best pages and cutting repetition. Exa says the highlights use 95% fewer tokens than full page text.
Fewer tokens, but a slower auto run
Artificial Analysis reported that Exa’s auto mode with Dynamic Highlights scored 80 on its Search Index, tying Perplexity Search (medium); Lite scored 79. The evaluator tested both Exa modes before release using its Stirrup agent harness and GPT-5.6 Luna (medium). Its index combines results from three question-answering benchmarks, so the tie describes this evaluation, not search performance in every setting.
In those tests, auto with Dynamic Highlights averaged 7.6 searches per task, down from 9.4 with standard highlights. Total model tokens fell 47%, from 299,000 to 157,000 per task, while cost fell 21%, from $0.083 to $0.066 per task. that task cost includes both search and model inference; it is not the price of a single API request. Exa lists auto at $7 per 1,000 requests and Lite at $4 per 1,000 requests.
