EN

Perplexity: an answer engine that builds its own Sonar models

Perplexity is an American company known for its answer engine that cites its sources. It developed its own models, Sonar, tuned for web search, while also giving access to models from other labs.

Data as of October 22, 2026

Key figures

CountryUnited States
HeadquartersSan Francisco
Founded2022
TypeStart-up
Weights policyMixed (open and closed)
Model familiesSonar
Consumer assistantPerplexity
Official siteperplexity.ai · Developer API
Models tracked here1

Perplexity was founded in 2022 in San Francisco. Its flagship product, the Perplexity assistant, answers questions in plain language and shows the sources it used next to each answer. This idea of AI-powered search is the heart of the company: rather than going head to head with the biggest general-purpose models, Perplexity set out to excel at one job, finding and summarizing information.

Sonar: models built for cited answers

To do that, Perplexity created Sonar, its in-house model family. The first modern Sonar, announced in February 2025, started from Meta’s open Llama 3.3 70B model, which Perplexity further trained to improve factual accuracy and readability. It is a good illustration of what open weights make possible: building a specialized product on top of a model released by someone else.

The family comes in several depths of research:

  • Sonar for simple questions and quick summaries;
  • Sonar Pro for everyday multi-step research;
  • Sonar Reasoning Pro for cross-checking many sources;
  • Sonar Deep Research for in-depth investigations.

Mostly closed, with a few open releases

Sonar models cannot be downloaded: you use them through the assistant or the API. Perplexity does, however, publish specialized open-weight models on Hugging Face, such as embedding models for semantic search. Its policy is therefore mixed, with a closed core product.

How to access Perplexity’s models

The developer offering has changed. According to the official documentation, support for the original Sonar Chat Completions interface ended on September 27, 2026; Perplexity now points developers to the Agent API, which combines Sonar, models from other labs and search tools. Other APIs round out the lineup: the Search API (raw search results), the Embeddings API and the Orchestrator API, which serves open-weight models hosted by Perplexity.

To compare Perplexity with other assistants, try our AI assistants comparison or take the quiz to find the right AI.

Perplexity (Sonar) models tracked on this site

ModelsReleasedCapability indexPrice (input / output)
Sonar Deep ResearchFebruary 1, 2025—$2 / $8

Frequently asked questions

Does Perplexity have its own AI models?

Yes. Sonar is Perplexity's in-house model, designed to produce factual, readable answers from search results. It was built on Meta's open Llama 3.3 70B model and then further trained by Perplexity.

Which models does the Perplexity assistant use?

Sonar is the default search model, but subscribers can also pick models from other labs such as OpenAI, Anthropic or Google. Perplexity is therefore both a model maker and a gateway to other models.

Is the Sonar API still available?

According to the official documentation, support for the Sonar Chat Completions interface ended on September 27, 2026. Perplexity recommends moving to the Agent API, which gives access to Sonar and third-party models with built-in web search.

Does Perplexity release open-weight models?

Yes, but not Sonar. Perplexity publishes specialized models on Hugging Face, including embedding models (pplx-embed) and classifiers. Sonar models are only available through its own services.

Data sources

Data as of October 22, 2026

Report an error

Does a figure, a date or a fact look wrong? Tell us: every report is checked.