سونار پرو

Perplexity's advanced search-grounded model, built for complex queries and follow-ups. It delivers deeper content understanding with more accurate search results, returning roughly twice the search results and twice the citations per search compared with standard Sonar. Perplexity calls it their best-performing model on factuality, leading the SimpleQA benchmark; every answer arrives with its sources attached.

Perplexity's case for Sonar Pro was about where an answer comes from rather than how clever the model is. Most generative features answer only from training data; the lab's position is that accuracy needs a live connection to the internet and answers drawn from sources a reader can check. Sonar Pro became generally available to every developer on the day of that post, alongside wider access to JSON mode and to search domain filters. The premise changes what is worth asking. A question about this week, about a price that moves, about a rule that changed - the kind a closed-book model can only guess at - is the kind this model was built to take, and what comes back carries the pages it was drawn from.

The evidence the lab put forward was SimpleQA, which asks short fact-seeking questions and scores whether the answer is right. Perplexity reported Sonar Pro leading that benchmark with an F-score of 0.858 against 0.773 for standard Sonar, and explained the result as an LLM's summarising combined with live retrieval instead of stored training data. That also explains the shape of the benchmark tiles above. Perplexity's own documentation calls Sonar Pro a non-reasoning model, and evaluations like GPQA Diamond and AIME are closed-book tests of precisely the recall and derivation this model is designed not to lean on. They are the cost of the trade, not a defect in it.

The uses the lab put forward all have that shape. Copy AI's go-to-market teams research prospects and target companies with it, which Perplexity says saved eight hours of research per representative per week and lifted throughput by a fifth. Doximity answers doctors' questions about guideline changes and insurance reimbursement, where the lab notes that in-line citations matter as much as the answer does. Zoom wired it into AI Companion 2.0 so someone can search without leaving a call. Sonar Pro also carries a larger context window than Sonar for longer, more nuanced searches and follow-up questions, and the lab's documentation states it does not train on customer data.

Two practical things follow from being a search product rather than a chat model. Answering is billed per request as well as per token, and that request fee rises with the search context size - Perplexity's settings run from a default tuned to be fastest and cheapest up to a maximum-depth setting meant for research - so how much of the web an answer is worth is a decision the caller makes and pays for. And the ceiling is real. On the lab's own migration benchmarks Sonar Pro scores 7.33% on BrowseComp against 16.33% for Sonar Reasoning Pro and 29.10% for Sonar Deep Research, with 30.11 F1 on DeepSearchQA and 17.92 on WideSearch. Perplexity now maps Sonar Pro workloads onto the low preset of its Agent API, saying the step up unlocks the genuinely multi-step agentic research Sonar Pro does not reach, and says the Sonar endpoint stays supported until 27 September 2026.