Perplexity API

Overview

Perplexity's API returns LLM answers grounded in live web search, with citations. Sonar Chat Completions (/chat/completions, models sonar, sonar-pro, sonar-reasoning-pro) reached end of support on 27 September 2026: sync and streaming calls are still being served by reformulating them as Agent API requests, but async requests no longer work. New code should use the Agent API (client.responses.create). sonar-reasoning and r1-1776 are deprecated.

Setup

pip install perplexityai          # Python
npm install @perplexity-ai/perplexity_ai   # TypeScript
export PERPLEXITY_API_KEY="pplx-..."       # read automatically by the SDK

Instructions

Basic query with web search

Pick a preset (tuned model + search depth) or a specific model.

from perplexity import Perplexity

client = Perplexity()  # uses PERPLEXITY_API_KEY
response = client.responses.create(
    preset="fast",
    input="What changed in the latest Python release?",
)
print(response.output_text)

Presets are dynamic: Perplexity tunes the model behind each name over time. Copy a preset's values inline if you need frozen behavior, and override any single parameter (model, tools, max_steps) while keeping the rest.

Choose a model, enable web search, add instructions

response = client.responses.create(
    model="openai/gpt-5.6-sol",
    input="Summarize this week's AI regulation news in the EU.",
    tools=[{"type": "web_search"}],
    instructions="Use web_search for current events. Keep queries brief. Cite sources.",
    max_output_tokens=2048,
)
print(response.output_text)

The Agent API also exposes models from Anthropic, Google and xAI under provider/model ids; check the models page for the current list.

Citations and search results

print(response.output_text)

for r in response.search_results:       # title, url, date, snippet
    print(r.title, r.url)

# Inline citation annotations live on the message content
for item in response.output:
    if item.type == "message":
        for part in item.content:
            for note in getattr(part, "annotations", []) or []:
                print(note)

Streaming and background mode

for event in client.responses.create(preset="fast", input="Top tech news today", stream=True):
    if event.type == "response.output_text.delta":
        print(event.delta, end="")

job = client.responses.create(preset="high", input="Full market analysis of EU battery makers", background=True)

Use background=True for long jobs and poll by response id; this replaces async Sonar requests.

Migrating old Sonar code

# before
client.chat.completions.create(model="sonar", messages=[{"role": "user", "content": q}])
# after
client.responses.create(preset="fast", input=q)

messages/choices become input/output; max_tokens becomes max_output_tokens, search_domain_filter and search_recency_filter move into the web_search tool's filters, and reasoning_effort becomes reasoning.effort. search_language_filter, stream_mode, video results and regex response_format have no equivalent. Read text from response.output_text instead of choices[0].message.content. For multi-turn chats pass earlier turns in input as the docs describe for conversation context.

TypeScript

import Perplexity from '@perplexity-ai/perplexity_ai';

const client = new Perplexity();
const response = await client.responses.create({ preset: 'low', input: 'Compare Postgres 17 and 18 release notes' });
console.log(response.output_text);

Examples

Example 1: Research assistant with sources

Request: "Answer 'which EU countries passed AI laws this year' and list the sources."

response = client.responses.create(preset="low", input="Which EU countries passed national AI laws this year?")
print(response.output_text)
for r in response.search_results:
    print("-", r.title, r.url)

Result: a cited paragraph, then one line per source page.

Example 2: Retry on rate limits

Request: "Make my lookup tolerant of 429s."

import time
import perplexity

def ask(q, retries=3):
    for attempt in range(retries):
        try:
            return client.responses.create(preset="fast", input=q).output_text
        except perplexity.RateLimitError:
            time.sleep(2 ** attempt)
    raise RuntimeError("rate limit retries exhausted")

Result: the call backs off 1s, 2s, 4s before giving up.

Guidelines

• Do not start new projects on /chat/completions or the openai SDK with base_url="https://api.perplexity.ai"; it works only through a compatibility layer being phased out.
• Search runs per request, so expect seconds of latency; fast is cheapest, high/xhigh cost far more.
• Tool calls (web search, URL fetch) are billed per call on top of tokens; check response.usage for cost.
• For exact prices or stock quotes, call a primary data API rather than trusting a summary.
• Keep the API key in the environment, never in source.
• Exact error class names are in the SDK reference; verify them there before relying on them. The REST endpoint is POST https://api.perplexity.ai/v1/agent.