This week, Perplexity announced its Sonar API's availability, enabling developers to integrate Perplexity's generative search capabilities into their custom applications. Perplexity has made a name for itself by combining LLMs with real-time internet search capabilities, developing an approach to generative AI applications that overcomes one of the main limitations of using plain LLMs.

While the latter have their capabilities substantially restricted to their training data, meaning their outputs can only be as high-quality and diverse as the data with which they were trained, Perplexity's products can increase not only the diversity but also the accuracy of their outputs. In particular, the enhanced accuracy of Perplexity's outputs is further backed by the inclusion of citations that enable users to trace any claims contained in an output to a reputable source.

The Sonar API will be available in two flavors, the upcoming basic Sonar API, designed to be faster and more affordable thanks to a flat pricing schema: every 1000 searches are $5, plus $1 per million input and $1 per million output tokens. In contrast, the Sonar Pro API (now available) aims to maximize accuracy (it provides twice as many citations as the basic Sonar API) and output quality, in addition to being able to handle more complex queries with a larger context window and the capability to handle follow up questions.

Perplexity claims that early adopters like Copy AI, Zoom and Doximity have already seen results. According to the announcement Sonar has helped Copy AI save approximately 8 hours of research time per representative weekly, leading to a 20% increase in throughput. In the healthcare sector, Doximity has implemented the technology to provide doctors with research-backed medical information, while Zoom's AI Companion 2.0 uses Sonar Pro to enable real-time search capabilities during video calls.

Perhaps most notably, Sonar Pro has achieved a leading F-score of 0.858 on the SimpleQA benchmark, a standard test for factual accuracy in language models, and is followed by Sonar, which scored 0.773.