Perplexity Sonar Deep Research is an agentic model built around the idea of research as an active loop rather than a single answer. Rather than drawing only on what was learned at training time, it autonomously plans a multi-step investigation: breaking a complex prompt into sub-questions, browsing the web, reading pages, cross-referencing findings, and refining its approach as evidence accumulates. That design intent shows up in its behavior, producing long, well-structured reports that resemble mini-papers with an introduction, sectioned findings, and a synthesized conclusion grounded in roughly ten to twenty cited sources per query. With a 128,000-token context window and a maximum output ceiling near 32,000 tokens, it can keep long research sessions coherent and carry substantial intermediate evidence through to a final write-up.
In practical use, Sonar Deep Research leans on its reasoning capabilities together with live web search to specialize in single-topic deep dives, where it can read and weigh dozens of sources before committing to an answer. It typically finishes most runs in under three minutes, which makes it the fastest of the major deep-research agents and the only one shipping through a pay-as-you-go research API rather than only as a product feature. The trade-off is that it tends to default to a full research pass even on simpler questions, so it shines most on tasks that genuinely benefit from broad synthesis: market scans, literature reviews, policy comparisons, and any investigation where pulling many threads into one readable report is the actual goal.
