Perplexity: Sonar Reasoning Pro

Perplexity: Sonar Reasoning Pro is a text generation model from Perplexity, released by the provider on 7 March 2025. It sits in the text category and is built as a premier reasoning model powered by DeepSeek R1 with Chain of Thought. If you already route other Perplexity models through your stack, this one comes from the same provider and follows the same general access pattern.

The model accepts text and image inputs and returns text. Its context window is 128,000 tokens, and the maximum output is 115,200 tokens, so long inputs and long answers are both viable within a single request. Because it is a reasoning model with Chain of Thought, it can expose its reasoning process, which you control through the reasoning and include_reasoning parameters. That combination of a large context, image input and explicit reasoning output makes it suited to tasks where the answer benefits from intermediate steps rather than a single direct response.

Practical uses follow from those traits. You can send a document, a screenshot or a diagram alongside a question and ask for an analysis that walks through the logic. Multi step problems, structured comparisons, and questions that require weighing several pieces of evidence are natural fits, since the model can reason before producing the final text. The web_search_options parameter also lets you attach search behaviour to a request, which helps when the answer depends on current information rather than only what is in the prompt. With a 128,000 token context you can supply large source material and still leave room for a detailed answer.

You call it through api14's OpenAI-compatible API using the model id "perplexity/sonar-reasoning-pro". The supported request parameters are frequency_penalty, include_reasoning, max_tokens, presence_penalty, reasoning, temperature, top_k, top_p and web_search_options. If you are porting existing OpenAI-style code, the request shape stays the same and you only change the model id and add the parameters you need.

Keep the output ceiling in mind: max_tokens can go up to 115,200, but reasoning tokens and the final answer both draw on that budget, so leave headroom when you want a long response. If your task is simple text generation with no need for Chain of Thought, image input or search, a lighter text model may be a better fit.