Perplexity: Sonar Pro Search

Perplexity: Sonar Pro Search is a text model from Perplexity, released by the provider on 30 October 2025. It is described as Perplexity's most advanced agentic search system, built for deeper reasoning and analysis, and it is exclusively available through the API rather than the consumer product. If you are already routing other models through api14, this is the Perplexity entry aimed at search heavy workloads rather than plain chat.

The model accepts text and image input and returns text output. It has a context window of 200,000 tokens and a maximum output of 8,000 tokens. Because it is an agentic search system, requests can be paired with web search options, and the model supports a reasoning mode as well as an option to include reasoning in the response. That combination makes it suited to questions where the answer depends on current information plus some analysis of what was retrieved, rather than a single pass over a static prompt.

Practical uses follow from those traits. Research assistants that need to gather sources and then compare or summarise them are a natural fit, as are competitive and market scans where the underlying facts change often. Image input extends this to tasks like reading a screenshot, chart or document page and then searching for related context. Analysis heavy workflows, such as pulling together findings across several topics into one reasoned answer, also map well onto the agentic search design. Straightforward generation tasks with no retrieval component are usually cheaper and simpler on a standard text model.

You call it through api14's OpenAI compatible API using the model id "perplexity/sonar-pro-search". The supported request parameters are frequency_penalty, include_reasoning, max_tokens, presence_penalty, reasoning, structured_outputs, temperature, top_k, top_p and web_search_options. In practice you would set web_search_options when you want the search behaviour engaged, use reasoning or include_reasoning when you need the model's reasoning surfaced, and rely on structured_outputs when the result has to be parsed by code. Sampling controls such as temperature, top_p and top_k, along with the penalty parameters and max_tokens, behave as they do for other OpenAI compatible endpoints.

The main constraints to keep in mind are the 8,000 token output ceiling and the text only output modality, so very long generated documents or non text outputs are out of scope. Input images are accepted, but the response is always text. If your task does not need search or multi step reasoning, a general purpose text model will likely be a better fit.