Perplexity AI Search Engine

Perplexity is an answer engine that queries the live web, retrieves the most relevant sources, and synthesizes a cited response using a large language model. It is not a chatbot you prime with a system prompt. It is a research tool: fast, sourced, and designed for questions that need current information rather than pre-trained knowledge. For technical teams, it replaces the workflow of “Google it, open five tabs, skim, summarize mentally.”

Official site: https://www.perplexity.ai
API documentation: https://docs.perplexity.ai
Pricing: https://www.perplexity.ai/pro

How it works

Perplexity is a RAG (retrieval-augmented generation) system built on top of live web search. When you submit a query, Perplexity runs a web search, retrieves the top-k most relevant pages, and feeds those pages into an LLM alongside your question. The model synthesizes a response and marks every factual claim with a numbered citation linking to its source.

The free tier answers with Perplexity’s own Sonar model. Pro subscribers can also pick third-party frontier models per query. This page has not re-verified the consumer app’s exact model picker, and the old Sonar Small / Large / Huge names it used to list are no longer current. The developer API’s Agent API does publish its catalog. As of 25 September 2026 it includes GPT-6 Sol and GPT-6 Luna, Claude Opus 5.5, Grok 4.7 and Gemini 3.8 Flash, all added in September, alongside GPT-5.6, Claude Sonnet 5 and Fable 5.1 (see OpenAI API and Claude Anthropic for the full current lineups).

This pipeline is distinct from a standard LLM chat session. The model’s knowledge cutoff is irrelevant: every response draws on pages retrieved seconds before the answer is written. That makes Perplexity reliable for current pricing, recent product launches, live API documentation, and breaking technical news.

Research workflow

Key features

Inline citations: Every factual claim in the response carries a numbered superscript. Click it to open the exact source page. This is the single most important difference from ChatGPT or Claude without web search. You can verify the claim in under five seconds.

Pro Search mode: Runs a deeper research pass. More sources retrieved, longer synthesis, more thorough cross-referencing. Use it for competitive intelligence, market sizing, or any question where the standard pass feels thin.

Focus modes: Narrow the search index before retrieval. Options include Web (default), Academic (peer-reviewed sources), YouTube (video transcripts), Reddit (community discussion), and Wolfram Alpha (mathematical computation). Academic mode is practical for literature reviews. Reddit mode surfaces practitioner opinions that do not appear in documentation.

Spaces: Shared research collections for teams. A Space is a persistent thread environment with its own system prompt, shared history, and access controls. Use it to run ongoing competitive monitoring or to build a shared research base for a client engagement.

File upload: Upload a PDF and ask questions against it. Perplexity combines document retrieval with web search in the same response. Useful for cross-referencing a vendor whitepaper against current market conditions.

API (Agent API, formerly Sonar): Perplexity exposes live web search to developers through an OpenAI-compatible REST API and returns cited responses in JSON. The Sonar Chat Completions API is being replaced by the Agent API. Covered in detail below.

Perplexity API

Deadline: Sonar Chat Completions is supported only until 27 September 2026. Perplexity’s docs say “Sonar Chat Completions is now Agent API” and recommend migrating existing Sonar usage and using the Agent API for all new projects. The Sonar example further down is kept for reference only. Start new work on the Agent API.

Agent API (current). The endpoint is POST https://api.perplexity.ai/v1/agent, with POST /v1/responses accepted as an alias for OpenAI SDK compatibility. It follows the Open Responses format: you send an input and get back a typed output array, with a message item for the answer and a search_results item for the sources, instead of messages and choices. You either name a model from any supported provider (OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI, NVIDIA, or Perplexity’s own Sonar) or call a preset. The presets are fast, low, medium, high, xhigh and wide-research, renamed from fast-search, pro-search, deep-research, advanced-deep-research and ultra. Built-in tools include web search, URL fetching, a code sandbox, MCP servers and finance and people search. September 2026 added GPT-6 Sol and Luna, Claude Opus 5.5, Grok 4.7 and Gemini 3.8 Flash to the catalog, plus a cheaper Fast Search mode at $1.00 per 1,000 web_search calls. Perplexity will also retire openai/gpt-5.4, gpt-5.4-mini, gpt-5.4-nano, gpt-5.2, gpt-5.1, gpt-5 and gpt-5-mini on 24 October 2026, so update any fallback chains that name them.

from perplexity import Perplexity

client = Perplexity()  # reads PERPLEXITY_API_KEY

response = client.responses.create(
    preset="low",  # everyday research with inline citations
    input="What replaced the OpenAI Assistants API after its 26 August 2026 shutdown?",
)

print(response.output_text)

Presets are not version-pinned. Perplexity updates the underlying model and configuration as its evals improve, so if you need reproducible behaviour, copy a preset’s current values into the request as a frozen configuration.

Sonar Chat Completions (legacy, supported until 27 September 2026). The original API gives access to the Sonar model family with live web search built in. It uses the OpenAI chat completions format.

Base URL: https://api.perplexity.ai

Available models:

sonar-reasoning no longer appears in Perplexity’s Sonar model list. Context figures are from this page’s June 2026 version and have not been re-checked.

Python example:

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["PERPLEXITY_API_KEY"],
    base_url="https://api.perplexity.ai",
)

response = client.chat.completions.create(
    model="sonar-pro",
    messages=[
        {
            "role": "system",
            "content": (
                "You are a technical research assistant. "
                "Return concise answers with all sources cited."
            ),
        },
        {
            "role": "user",
            "content": (
                "What is the current pricing for Amazon Bedrock Claude Sonnet "
                "in eu-west-1 as of June 2026?"
            ),
        },
    ],
)

answer = response.choices[0].message.content
citations = response.citations  # list of URLs used in the response

print(answer)
print("\nSources:")
for i, url in enumerate(citations, start=1):
    print(f"  [{i}] {url}")

The response.citations field returns the list of URLs that the model used to construct the answer. Store these alongside any extracted facts if you are building a research pipeline that requires audit trails.

Cost: Sonar is priced per token plus a per-request fee that depends on search context size. As of 26 September 2026: sonar costs $1 input / $1 output per million tokens plus $5 / $8 / $12 per 1,000 requests (low / medium / high context); sonar-pro costs $3 / $15 per million tokens plus $6 / $10 / $14 per 1,000 requests; sonar-reasoning-pro costs $2 / $8 per million tokens with the same request fees as Pro; sonar-deep-research costs $2 / $8 plus $2 per million citation tokens, $5 per 1,000 searches and $3 per million reasoning tokens. Remember that Sonar Chat Completions is supported only until 27 September 2026. The Agent API instead bills model tokens at the per-model rates on its models page, plus a per-call fee for tools such as web_search. Check the official pricing page for current rates: https://docs.perplexity.ai/docs/getting-started/pricing

Comparison

When to use Perplexity

Technical research: Ask “What replaced the OpenAI Assistants API after it shut down on 26 August 2026?” and get a sourced answer in fifteen seconds. No tab-opening required.

Competitive intelligence: “What are the key differences between Snowflake Cortex and Amazon Bedrock Knowledge Bases?” returns a comparison synthesized from current vendor documentation, analyst posts, and community discussion.

Documentation lookup: Use it for current pricing, quota limits, and feature availability. Official docs change frequently. Perplexity retrieves the live page, not a cached version from training data.

Replacing long-form Google sessions: When a question requires synthesizing five or more sources, Perplexity does the synthesis. Use this for market sizing, vendor shortlisting, or technology selection research.

Academic focus mode: Surface peer-reviewed papers on a topic without using Google Scholar directly. Useful as a starting point before going deeper into a literature review.

When not to use Perplexity

When you need to verify every claim independently: Perplexity is a synthesis tool. It can misread or misrepresent sources. For high-stakes decisions, open the citations and read the originals yourself. Do not treat the synthesized answer as a primary source.

When you need deep analysis, not synthesis: Perplexity is fast and broad. For nuanced interpretation of a complex topic (for example, a detailed code review or architectural trade-off analysis), a long-context LLM with your own curated context will outperform it.

When API cost matters for non-search tasks: Sonar queries the web on every call, which adds latency and cost compared to a standard LLM completion. If your task does not require live web data (classification, summarization of a document you already have, code generation), use Claude or GPT-5.6 directly. The web search overhead is unnecessary.

When you need reproducible outputs: Web search results change. The same query submitted twice may produce a different answer if the top-k retrieved pages differ. For deterministic pipelines, use a static knowledge base.

When source quality needs strict control: Perplexity retrieves from the open web. The Academic focus mode narrows this, but the default mode may cite blog posts, forums, or vendor marketing. For regulated industries (healthcare, finance, legal), validate every source before relying on the output.

Further reading

• Perplexity API documentation : Official reference for the Agent API, Search API, Router API and legacy Sonar API.
• Perplexity API pricing : Current model, request and tool rates.
• Sonar models : sonar, sonar-pro, sonar-reasoning-pro and sonar-deep-research, with the 27 September 2026 support deadline.
• Perplexity Spaces documentation : How to create shared research collections and configure custom system prompts for team use.
• What is RAG? : Explains retrieval-augmented generation, the architecture that underlies Perplexity’s answer pipeline.
• What is an LLM? : Background on large language models, which Perplexity uses for synthesis.
• LLM Landscape 2026 : Model comparison covering the Sonar family alongside GPT-5.6, Claude, and Gemini.
• Perplexity research hub : Blog posts and product announcements from the Perplexity team.

Sources

1. Perplexity, “Migrate from Sonar to the Agent API” (Sonar supported until 27 September 2026; Sonar vs Agent API capabilities), checked 25 September 2026: https://docs.perplexity.ai/docs/agent-api/migrate-from-sonar/overview
2. Perplexity, Agent API quickstart (/v1/agent endpoint, /v1/responses alias): https://docs.perplexity.ai/docs/agent-api/quickstart
3. Perplexity, Agent API models: https://docs.perplexity.ai/docs/agent-api/models
4. Perplexity, Presets (tier-based preset names, dynamic vs frozen presets): https://docs.perplexity.ai/docs/agent-api/presets
5. Perplexity, API changelog (September 2026: GPT-6 Sol/Luna, Claude Opus 5.5, Grok 4.7, Gemini 3.8 Flash, Fast Search, 24 October 2026 OpenAI model retirements): https://docs.perplexity.ai/docs/resources/changelog
6. Perplexity, Sonar models: https://docs.perplexity.ai/docs/sonar/models
7. OpenAI, Deprecations (Assistants API shut down 26 August 2026): https://developers.openai.com/api/docs/deprecations
8. Google for Developers, Custom Search JSON API overview (closed to new customers; discontinued 1 January 2027): https://developers.google.com/custom-search/v1/overview