Sonar by Perplexity: The Fastest AI Search Model for Accurate, Real-Time Answers

Launched in 2022, the Perplexity AI search engine received a significant update earlier this year. The developers presented the audience with the long-awaited Sonar API and a series of language models of the same name, available for integration with external systems. From our article, you will learn about what Perplexity Sonar is, what capabilities its AI models have, how much they cost and where they can be used, as well as how they are superior or inferior to competitors.

Sonar is a series of language models (LLM) and the API of the same name for a search engine with built-in artificial intelligence Perplexity AI. The solution is based on the open-source model Llama 3.3 70B from Meta, and its official release took place in February 2025. After the appearance of Sonar AI, the use of Perplexity AI tools became possible not only through the web interface and mobile application, now the system supports integration with third-party services and applications via API.

Source: sonar.perplexity.ai

One of the first Sonar users was Zoom, which integrated its models into the AI assistant of its communications platform. The developers released several products at once, including the flagship Sonar Pro model and its lightweight version, Sonar. In addition, the line includes several LLMs for solving complex problems, Sonar Reasoning and Sonar Reasoning Pro, as well as a model for expert research, Sonar Deep Research.

A special place in the Sonar Perplexity AI family of models is occupied by the R1 1776, a special version of the LLM DeepSeek R1, trained to provide unbiased factual information without censorship. We will tell you more about these AI tools in the next section of the article.

Now that you know what Sonar in Perplexity AI is, it's worth talking about the underlying technology. The line of models is based on the open-source LLM Llama 3.3 70B, released by Meta Corporation in December 2024. It has 70 billion parameters and is designed to solve a wide range of text-related tasks: multilingual chat with support for 8 languages, code processing, synthetic data generation, etc.

Llama 3.3 is not a multimodal language model, so it cannot handle visual or audio content. Its main advantage is that this LLM is optimized to run on regular GPUs, so it is well suited for deployment on local devices. It also demonstrates decent performance, requiring fewer resources compared to more powerful analogues.

The Sonar line includes the following models:

The full-size version of LLM has an enlarged context window, which gives it additional memory and increases efficiency. The Pro version can generate detailed citation responses, with twice as many links as other models in the family. It is ideal for solving complex multi-stage problems that require deep understanding and context preservation. Sonar Pro (like the rest of the LLM series) does not use user queries or other data for training, ensuring complete confidentiality.

The junior version of the model is trained on a smaller dataset, which provides it with an optimal balance between performance, cost, and query processing speed. The key advantages of Sonar are accelerated response generation and more flexible pricing, compared to the full-size Sonar Pro. This LLM provides accurate and capacious responses with citations, quickly finding information on the Internet in real time.

The next of the Perplexity Sonar models has improved planning and reasoning capabilities, can create logical chains and search the web for factual, unbiased information in real time. Sonar Reasoning Pro generates detailed, highly cited answers and effectively combines data from multiple sources. It is based on the open-source LLM DeepSeek R1, optimized for analyzing multiple types of data and solving complex data-processing problems.

The junior version of the Sonar Reasoning Pro model performs planning and reasoning queries at an accelerated rate, and also selects necessary citations and searches for facts on the Internet. The LLM series does not censor the data it processes and generates, and the equipment that serves them is located in data centers in the United States.

A specialized model in the Perplexity Sonar family, it is designed for deep research on narrow topics. It runs dozens of search queries and processes hundreds of sources, and can reason and generate detailed insights on a wide range of topics — from marketing and finance to technology and tourism. Once Sonar Deep Research has analyzed the raw data, it produces a detailed report with expert opinion.

An open-source model developed on DeepSeek R1 and post-trained to provide the most unbiased and accurate factual information possible.

In this section, you will learn about Perplexity Sonar API pricing, which includes the following plans:

Most Sonar models support three operating modes, which differ in a number of parameters. The High mode provides the maximum depth of research and preserves context for processing the most complex queries. The Medium mode is balanced in terms of performance and cost — it is advisable to use it for solving moderately complex problems. The Low mode is the most economical and at the same time maintains high accuracy when processing simple queries.

Users can track expenses for each of the API keys they've added. To achieve this, go to the API section in your account settings, then select Usage Metrics > Invoice History > Invoices. Then click on any of the invoices available there to view expense details for the selected time period.

The Perplexity Sonar API and its family of models have a very wide range of applications, including:

A/B testing conducted by the developers showed that the top model of the Sonar family outperforms its main competitors in its class, in particular GPT-4o mini and Claude 3.5 Haiku. It outperformed its competitors in such parameters as readability and factual accuracy. The results of Sonar Pro correspond to the leaders in its segment, LLM GPT-4o and Claude 3.7 Sonnet.

The impressive performance of the Sonar Pro, the most powerful model in the line, is largely due to its Cerebras inference infrastructure. Thanks to it, Sonar by Perplexity AI generates responses at an incredibly high speed of 1200 tokens per second — almost 10 times faster than the results of the competing model Gemini 2.0 Flash.

When it comes to the price factor, Sonar again outperforms its closest competitors, as its models are less expensive for users. For example, using the top version of Sonar Pro via the API costs $3/$15 per million I/O tokens. Meanwhile, OpenAI’s GPT-4o costs $5/$20 per million I/O tokens when solving text-based problems in real time. And Anthropic’s Claude 3 Opus, which is designed for complex queries, is even more expensive at $15/$75 per million I/O tokens.

Perplexity Sonar AI is a game changer in the field of search AI, as its API and state-of-the-art language models open up broad opportunities for automated data collection and analysis across third-party systems and applications. The platform offers users a wide range of LLMs with different performance, speed, and price ranges, allowing to choose the best option for a specific area and task. And its unique combination of high-speed real-time web search, deep analytics capabilities, and versatility make Sonar a potential new leader in the highly competitive AI industry.