Perplexity Sonar Pro

Acerca de Perplexity AI

Fundado en 2022 · San Francisco, USA

Perplexity AI was founded in August 2022 by Aravind Srinivas (CEO, ex-OpenAI and DeepMind/Google Brain), Denis Yarats (CTO, ex-Meta AI), Andy Konwinski (co-founder of Databricks) and Johnny Ho. The company built an answer engine that combines live web search with LLMs to return cited answers, positioning itself as a search alternative to Google. Perplexity rapidly raised over $500M from IVP, NEA, NVIDIA, Jeff Bezos, Susan Wojcicki and others, reaching a $9B valuation in late 2024 and a higher valuation in 2025. The Sonar model family debuted in late 2024 as Perplexity's in-house RAG-tuned models, with Sonar as the fast default tier and Sonar Pro as the higher-capability variant for advanced research workloads. Sonar Pro pairs deeper search (more queries per request, more sources synthesised) with a stronger underlying LLM, available via the Perplexity API and the Perplexity Pro consumer tier. The company also partners with Apple Intelligence and several US publishers for revenue-sharing on cited content. Beyond Sonar, Perplexity ships Perplexity Pages, Spaces and Discover surfaces, and runs ad-free Pro and Enterprise subscriptions.

Arquitectura

Search-grounded LLM (Llama-derived large, deeper retrieval pipeline)

Perplexity Sonar Pro is the high-capability tier of Perplexity's in-house Sonar family, launched alongside the Sonar API in January 2025. It is a Llama-3.x-derived large model fine-tuned by Perplexity for deep retrieval-augmented answering. Compared to base Sonar, Sonar Pro runs a more aggressive retrieval pipeline: it issues multiple parallel search queries per request, retrieves more documents (up to a few dozen sources), performs intermediate ranking and synthesis steps, and produces longer, more structured answers with denser inline citations. The underlying LLM is a larger and more capable checkpoint than base Sonar, trained with extended supervised fine-tuning on Perplexity's curated multi-hop research answers and preference optimisation against expert-rated grounding quality. Sonar Pro supports a 200K context window, structured output, and exposes additional response fields including a list of all sources visited and the search queries used. The API is OpenAI-compatible, making it a drop-in replacement for stacks that need higher answer quality and citation coverage. Pricing is roughly $5 per 1,000 search-augmented requests plus token costs, positioning it as a premium tier for research and enterprise use cases.

Parámetros

Undisclosed (Llama-3.x-derived large)

Contexto

200,000 tokens

Capacidades

• Deeper retrieval pipeline: multiple parallel searches and intermediate ranking
• 200K context window
• Denser inline citations and longer structured answers
• Exposes search queries and visited sources in API response
• OpenAI-compatible chat completions API
• Larger underlying LLM than base Sonar
• Suitable for multi-hop research questions
• Structured output (JSON schema) supported
• Available in Perplexity API and Perplexity Pro consumer tier
• Higher accuracy than base Sonar on long-form factual queries
• Best for: research workflows, financial and policy briefings, multi-hop fact-finding, premium chat experiences.

Entrenamiento y licencia

Built on a Llama-3.x-derived large base, fine-tuned on Perplexity's curated multi-hop research answers, expert-rated grounding preference data and synthetic deep-retrieval QA pairs. At inference time, augmented with multi-query retrieval from Perplexity's in-house search index.

Licencia: Proprietary commercial license via the Perplexity API and Pro/Enterprise subscriptions. Weights not released.

Pruebas de seguridad: Limited public safety evaluations. Refuses to answer when retrieved sources do not support a claim. Inherits some safety properties from the Llama base.