Sonar Reasoning thinks out loud before answering, working through problems step by step rather than jumping straight to conclusions. It's built on top of DeepSeek R1 and layered with Perplexity's live web search, so its reasoning is grounded in current information rather than frozen training data. The trade-off is that responses can be longer and slower than a quick-answer model.