Sonar Pro Search

Perplexity's most advanced agentic search system, enhancing Sonar Pro with automated tool use and multi-step reasoning through intelligent tool orchestration. Rather than running one query and synthesising the result, it plans and executes an entire research workflow, issuing follow-up searches and reading further before it answers. It is slower and dearer than plain Sonar Pro, and correspondingly better on questions that need real research.

Pro Search is not a separate model, and it is worth being plain about that: it is Sonar Pro called with its search type set to pro. The weights, the 200K context window and the per-token rates are Sonar Pro's, unchanged. What the setting changes is what the model is allowed to do before it answers.

In its ordinary mode Sonar Pro runs one search and synthesises what comes back. In pro mode it drives two built-in tools on its own - a web search, and a fetch that pulls a page's full contents rather than the snippet a result list gives - and it loops: search, read, search again, until it has enough to answer. Perplexity's worked example is a question about solar panels, where the model searches for current incentives and costs, opens the policy documents themselves, then searches again to check electricity rates against providers. The tool set is fixed and the model decides when to use each; a caller cannot register tools of their own.

Two constraints come with the mode. It exists only inside a streaming call - the lab's documentation says a non-streaming request falls back to ordinary Sonar Pro behaviour rather than doing the extra work - and it is dearer per request rather than per token, with Perplexity's rate card leaving the input and output rates alone and raising the search request fee to more than double the standard Sonar Pro fee at the default search context size. Because paying that on a single-fact lookup is waste, the lab also ships a classifier: set the search type to auto and each query is routed to pro or to fast on its own complexity, which Perplexity recommends as the approach for most applications. Every tool call the model makes streams back as a reasoning step, so the research is visible while it happens instead of summarised once it is over.

The benchmark tiles above measure the underlying model, and the underlying model is Sonar Pro - the search mode changes what reaches the context, not what the weights know. That is the right way to read them here. Pro Search does not make Sonar Pro a better reasoner; it makes the evidence in front of the reasoning larger, fresher and read at full length, which is the half of a research answer a closed-book evaluation does not measure. Perplexity is candid about the ceiling too: it maps Sonar Pro onto the low preset of its Agent API, reserving multi-hop browsing and wide aggregation across many sources for the tiers above it, and says the Sonar endpoint stays supported until 27 September 2026.