Exa 3.0 adds cross-result highlights to its search API

• Search excerpts now share a budget across results: Exa says its model gives stronger pages more space and cuts repeated information.
• Independent search quality improved: Auto with Dynamic Highlights rose from 74 to 80 on the Search Index, tying Perplexity, in Artificial Analysis's evaluation.
• Auto tasks got cheaper and slower: Artificial Analysis measured 21% lower combined cost but 48% longer completion time than with standard highlights.
• Lite search costs $4 per 1,000 requests, with Exa positioning it below the higher-compute Auto tier.

Exa's launch post puts its independent index at more than 2 trillion tracked URLs and 100 billion served pages. The benchmark harness gives agents 25 turns, with unlimited tool calls within that budget.

Exa's neatest change is giving search results a shared excerpt budget:

• Stronger pages get more of the available text budget.
• Repeated information across pages gets cut.

Exa claims 95% fewer returned tokens than full-page text. The retrieval stack also gets Exa's largest embedding model, trained with more data and compute than its predecessors.

Exa claims first place on every benchmark it tested, including coding and go-to-market intelligence. Independent testing produced these Artificial Analysis Search Index results:

| Search configuration | Index score |
| --- | ---: |
| Exa Auto, standard highlights | 74 |
| Exa Auto, Dynamic Highlights | 80 |
| Exa Lite, Dynamic Highlights | 79 |
| Perplexity Search, medium | 80 |

On DeepSearchQA, both new Exa configurations beat Perplexity Search medium's 80.8 F1:

• Auto: 85.6 F1, +4.8 points.
• Lite: 84.9 F1, +4.1 points.

The tests used GPT-5.6 Luna at medium effort in the Stirrup harness. The methodology defines the index as an equal-weighted average of three benchmarks and filters known or potential benchmark leaks from search results and fetched pages.

Artificial Analysis counts input, reasoning, and output tokens across the whole task, its evaluation explains. Its comparison uses standard highlights as the baseline; Exa's 95% claim uses full-page text.

| Metric | Auto, standard highlights | Auto, Dynamic Highlights | Change |
| --- | ---: | ---: | ---: |
| Searches per task | 9.4 | 7.6 | About 19% fewer |
| Total tokens per task | 299k | 157k | 47% fewer |
| Model inference cost per task | $0.018 | $0.012 | About 33% lower |
| Search plus inference cost per task | $0.083 | $0.066 | 21% lower |

Higher per-search latency outweighed the reduction in search calls for Auto. Artificial Analysis's task-time metric includes both search and model inference.

Both tiers query the same index, with different compute allocations:

| Tier | Price per 1,000 requests | Retrieval compute |
| --- | ---: | --- |
| Lite | $4 | Lower-cost configuration |
| Auto | $7 | More compute for candidate generation, highlights, and ranking |

Lite's combined search and inference cost was $0.044 per task in the independent evaluation, comprising approximately $0.031 for search and $0.013 for model inference.

Exa lists four coverage expansions:

• 1.27 billion ecommerce listings, including prices and availability.
• 211 million-plus businesses, including hours and ratings.
• Real-time news, sports, weather, and market data.
• Flights, local events, transit, and podcasts.

Exa announced API availability for the upgraded search, Lite, highlights, and expanded index on October 19.

Artificial Analysis's published request settings separate the search tier from excerpt selection:

• Search tier: type: "auto" or type: "lite".
• Dynamic Highlights: contents: {"highlights": {"dynamic": true}}.
• Standard highlights: contents: {"highlights": true}.

The launch post calls Dynamic Highlights generally available, but the Highlights guide still labels it a research preview. Its examples send Exa-Beta: dynamic-highlights-2026-08-28, or pass DYNAMIC_HIGHLIGHTS_BETA through the Python and JavaScript SDKs.