Perplexity Review

Perplexity is not a model, it is a router with a search index bolted on, and grading it means grading the wrapper. Underneath, queries go to GPT-5.5, Claude Opus 4.7, Gemini 3.1 Pro, Kimi K2.6, or its own Llama-derived Sonar 2, and you generally cannot tell which one answered you. That opacity is the core problem: the same question asked twice can land on different models with different failure modes, and there is no way to hold the output to a standard.

Citations look rigorous and frequently are not, pointing at pages that merely mention a term rather than support the claim being made. Answers are synthesized short by design, which caps depth on anything needing sustained reasoning, and it does not do agentic coding in any meaningful sense. What it gets right is the interface. It is fast, it is clean, and the cited-answer format genuinely beats a page of blue links for quick factual lookups at $20 a month. As a search product it works. As an AI you reason with, it does not.