AI search ranking: The AI models from Perplexity Sonar are leading in the AI search landscape

Published on: April 15, 2025 / update from: April 15, 2025 - Author: Konrad Wolfenstein

AI search ranking: The AI models from Perplexity Sonar are leading in the AI search landscape-Image: Xpert.digital

Sonar-Reasoning-Pro-High: Perplexity's jump to the top of the AI search

Change in AI search system: Perplexity's milestone in development

Perplexity's Sonar models have achieved impressive results in the latest LM Search Arena Evaluation, whereby Sonar-Reasoning-Pro-High after Google's Gemini-2.5-Grounding is a leader. This assessment represents an important milestone in the evolution of AI search systems and underlines perplexity's leading position in this competitive area.

Suitable for:

Perplexity Sonar Pro API as an AI search engine in external applications and tools – for smart apps and tailored search

The LM Search Arena Evaluation

The LM Search Arena is a novel evaluation platform developed by LM Arena to evaluate search-strengthened AI systems based on human preferences. In contrast to previous benchmarks such as Simpleqa, which concentrated on close factual accuracy, the Search Arena evaluates how models cut off for real user inquiries in areas such as programming, writing, research and recommendations.

The evaluation took place between March 18 and April 13, 2025 and collected over 10,000 human preference votes for 11 models. Users were asked to ask inquiries and then evaluate which model response their information needs better met.

Outstanding performance of the Sonar models

Perplexity's Sonar-Reasoning-Pro-High reached an arena score of 1136 (± 21/−19), which is statistically equivalent with Google's Gemini-2.5-process (1142 +14/-17) and thus means a common top position. It is particularly noteworthy that with direct comparisons Sonar-Reasoning-Pro-High Gemini-2.5-Pro-Grounding exceeded in 53% of cases.

The dominance of perplexity in the evaluation is illustrated by the following ranking:

Gemini-2.5-Pro-Grounding (1142 points)
Sonar-Reasoning-Pro-High (1136 points)
Sonar-Reasoning (1097 points)
Sonar (1072 points)
Sonar-Pro-High (1071 points)
Sonar-Pro (1066 points)

All perplexity models took the upper ranks and exceeded significantly different rated models from Google (Gemini 2.0-flash grounding) and Openai (GPT-4O Search).

Key factors for success

The Search Arena identified three factors that correlated strongly with human preference:

Superiority in sources

A higher number of quotations correlated strongly with the user preference (coefficient 0.234, p <0.05). The Sonar models carry out a deeper search and quote an average of 2-3 times more sources than comparable Gemini models. This comprehensive source use ensures that the information provided is well documented and trustworthy.

Use of various sources

The evaluation showed that quotes from community web sources were particularly valued. The Sonar models are characterized by the effective use of different sources, including YouTube, community platforms and authoritative sources.

Control experiments confirmed this findings and showed that the search depth is an essential difference in performance between the models. When checked for quotes, the model rankings converged, indicating that the search depth is a decisive differentiation factor.

Suitable for:

AI integration of an independent and cross-data source-wide AI platform for all company matters

The technology behind Sonar

Perplexity's Sonar model is based on Llama 3.3 70b and was specifically developed for the optimization of the answer quality and user experience. It was trained to improve the fidelity and readability of answers.

Speed and performance

Sonar is driven by the cerebras infrastructure and provides answers at impressive speed-1200 token per second, which enables almost immediate response generation. This speed is almost 10 times faster than with comparable models like Gemini 2.0 Flash.

User preference and performance comparison

Extensive A/B tests showed that sonar clearly exceeds models such as GPT-4O Mini and Claude 3.5 Haiku and even achieves the performance of top models such as GPT-4O and Claude 3.5 Bonnet when it comes to user satisfaction.

Sonar API: Accessibility for developers

Perplexity also offers its sonar technology via APIs, which enables developers to integrate AI-based search functions into their applications. There are two main versions of the API:

Sonar api

The standard sonar API is lightweight, inexpensive, quick and easy to use. It was designed for companies that need uncomplicated question-answer functions and are optimized for speed.

Sonar Pro API

For companies that need more advanced functions, the Sonar Pro API offers the opportunity to process more complex, multi -stage inquiries. On average, it generates twice as many sources per search as the standard version and has a larger context window for longer and more nuanced search queries.

The price structure reflects these differences: Standard sonar costs $ 5 per 1,000 plus $ 1 per 750,000 words (input and output combined). Sonar Pro keeps the same 5 $ 1,000 searches, but calculates $ 750,000 input words and $ 15 per $ 750,000 generated words.

From factors of accuracy to user orientation: Perplexity's sonar convinced

The outstanding results in the LM Search Arena Evaluation confirm that Perplexity's Sonar models are among the leading AI search systems. With the combination of fidelity, extensive source information and deep search ability, they offer a superior user experience.

These successes underline Perplexity's position as an innovator in the field of AI-based search and provision of information. The continuous improvement of the models based on user feedback indicates further potential for future developments.

For perplexity users, these results mean that they have access to first-class accuracy, extensive source attribution and high-quality answers to a wide range of topics. Pro users can continue to benefit from these powerful models by determining Sonar as their standard model in the settings.

The strong performance of Sonar in the Search Arena Evaluation not only underlines the technological competence of perplexity, but also shows the way for the future of looking for AI: more precisely, more comprehensive and with a deeper understanding of the information needs of the users.

Suitable for: