Exa claims Exa 3.0 search API leads Perplexity, Brave and Parallel on eight benchmarks
Exa Labs announced Exa 3.0, an upgraded web‑search API for AI agents. The release adds a larger, undisclosed embedding model, a new "Lite" search type priced at $4 per 1,000 requests, and Dynamic Highlights that cut token usage by 95% with no extra charge. Exa says its index now covers over 2 trillion URLs and serves 100 billion pages, including 1.27 billion ecommerce listings and 211 million…
Key points
- Exa Lite costs $4 per 1,000 requests; Auto costs $7 per 1,000 requests.
- Exa 3.0 Auto scores 91.8% on BrowseComp and leads on eight self‑run benchmarks, per Exa.
- Dynamic Highlights reduces token usage by 95% with no extra charge.
According to Exa’s self‑run evaluation, Exa 3.0 Auto ranked first on all eight benchmarks against Perplexity, Brave and Parallel. Scores include 91.8% on BrowseComp and a 42.0% lead on ATLAS (vs 37.2% for Perplexity). The biggest gaps are +5.6 points on WebCode and +4.8 on ATLAS. Cost‑per‑task figures show Lite scoring about 79 at $0.045 per task, while Auto scores about 80.5 at $0.065. Exa notes that Auto costs $7 per 1,000 requests, higher than the $5 charged by the three rivals.
Exa cautions that the benchmarks were run with its own harness and that some margins are thin. Lite rejects unsupported parameters with a 400 error, and the embedding model’s size and architecture remain undisclosed.
Model page: GPT-6 Luna →
The story so far
2 episodes →- Exa claims Exa 3.0 search API leads Perplexity, Brave and Parallel on eight benchmarksthis story
Exa 3.0 Search API Tops Perplexity, Brave and Parallel on 8 Benchmarks
MarkTechPost · 20 October 2026
Loading the full article…
This text was published by MarkTechPost and written by Michal Sutter. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Agents & Tools
All →- DeepSeek releases DeepInfer open-source inference framework for DeepSeek models · 1 src
- U.S. Army develops echotribble program to deploy autonomous AI agents for cyber attacks · 1 src
- Ready Server adds free ReadyServer LLM API to every VPS plan · 1 src
- In an era where we ask AI to 'do it for me,' the most important thing might be how we hold the reins—Reflections on the Gemini call agent and the OpenAI agent i · 1 src
- Researchers introduce Mirror, a self‑supervised video‑to‑action model · 1 src
Comments
via GitHub Discussions