Perplexity API pricing and model catalog: use-case guide

Unlike traditional LLM providers, Perplexity's Sonar models combine language generation with real-time web search. Every API call retrieves relevant web pages, synthesizes an answer, and returns inline citations with source URLs. This eliminates the need to build your own RAG pipeline for web-grounded Q&A.

Perplexity also offers an Agent API with tool-use capabilities (web search, URL fetch, people search, finance search) and a sandbox execution environment. This makes it suitable for multi-step research workflows beyond simple question answering.

The catalog also includes third-party models accessed through Perplexity's routing layer: Anthropic Claude, OpenAI GPT, Google Gemini, and xAI Grok variants, all with web search enabled.

Choose the right Sonar tier. The base Sonar model handles most search-augmented Q&A tasks. Reserve Sonar Pro and Deep Research for queries that genuinely require multi-step reasoning or exhaustive citation. See the model routing cascade framework.

Minimize search context depth. Sonar request fees vary by search context size (low/medium/high). For straightforward factual queries, the low tier is sufficient and costs less per request.

Use budget models for classification. If you only need to classify or route queries without search augmentation, use the cheaper Llama or Mistral models at $0.07–$0.20/1M input tokens instead of Sonar.

Batch embedding jobs. Perplexity's embedding models are priced per token with no per-request overhead. Process documents in large batches to amortize any fixed costs. See the embedding cost guide.

Consider the Agent API for complex workflows. The Agent API's preset tiers (fast, low, medium, high) bundle model, tools, and sandbox costs. For multi-tool workflows, a preset may be cheaper than composing individual API calls.

Does Perplexity charge for search separately? Yes. Sonar models charge a per-request search fee in addition to token costs. The fee varies by search context depth (low, medium, high). Agent API charges per-invocation tool fees for web_search, fetch_url, and other tools.

What is the difference between Sonar and Sonar Pro? Sonar is optimized for straightforward Q&A with search. Sonar Pro handles multi-step queries with deeper search context and higher-quality synthesis, but costs more per token and per request.

Does Perplexity offer cached input pricing? Yes. Both Sonar and Agent API models support cache pricing. Cached input tokens are billed at a fraction of the standard input rate, reducing costs for prompts with repeated prefixes.

Can I use Perplexity models without search? Some models in the catalog (Llama, Mistral, CodeLlama) are available as standard chat models without search augmentation. The Sonar models always include web search.

How does Perplexity compare to building my own RAG? Perplexity's Sonar models eliminate the need to build search infrastructure, manage vector stores, and handle citation extraction. For web-grounded Q&A, it can be simpler and cheaper than a self-built RAG pipeline. See the RAG application guide for comparison.

What embedding models does Perplexity offer? Two models: pplx-embed-v1-0.6b (1024 dimensions, $0.004/1M tokens) and pplx-embed-v1-4b (2560 dimensions, $0.03/1M tokens). Both support 32k context windows.