Perplexity: Sonar Reasoning

Perplexity's Sonar Reasoning model, based on DeepSeek R1, demonstrates exceptional reliability with a 98% success rate, consistently providing usable responses. However, it tends to have longer response times, ranking in the 18th percentile for speed, and is positioned at premium pricing levels, ranking in the 9th percentile for cost. The model exhibits a key strength in its ability to acknowledge uncertainty, achieving perfect 100% accuracy in the Hallucinations (Baseline) benchmark, indicating no hallucination. It also performs very strongly in Reasoning (98% accuracy, 92nd percentile) and Email Classification (99% accuracy, 72nd percentile). Its General Knowledge (97% accuracy) and Mathematics (89% accuracy) capabilities are solid. A notable weakness is observed in Instruction Following, where it achieved only 34.7% accuracy, placing it in the 35th percentile. Coding performance is moderate at 81% accuracy, and Ethics performance is 95% accurate but ranks in the lower 28th percentile. Its built-in web search and uncensored nature are additional features.