Responses API Quickstart

The Parallel Responses API is an OpenAI Responses-compatible endpoint that answers a question using live web research within a latency budget of ~5–60 seconds. The API automatically handles running multi-step web searches, fetching live pages, and cross-checking sources to return a synthesized answer with citations. Because it is OpenAI compatible, you can invoke the Responses API with OpenAI’s Typescript or Python SDKs. Just change three parameters: base_url, api_key, and model.

Model and reasoning tiers

There is one model id — parallel — and reasoning.effort selects the reasoning tier:

`reasoning.effort`	Latency	Designed for	Example
`low`	~5–10s	Simple fact-retrieval questions	”What’s the current U.S. federal funds rate?”
`medium` (default)	~15–20s	More complex, multi-hop fact-retrieval questions	”Which G7 country has the highest corporate tax rate?”
`high`	~30–60s	Deep-research questions requiring extensive search and synthesis	”How have G7 corporate tax rates shifted since 2020, and what’s driving the trend?”

Every tier grounds its answer in live web search, but higher tiers do more web research before producing an answer. If you omit reasoning, effort defaults to medium. high requests can take up to a minute, so give your client timeout headroom (e.g. timeout=120.0 in the Python SDK).

Prerequisites

Get an API key from Parallel Settings and set it as PARALLEL_API_KEY. If you are using an SDK, install the OpenAI SDK:

export PARALLEL_API_KEY="your-api-key"

pip install openai

npm install openai

Make your first request

curl https://api.parallel.ai/v1/responses \
  -H "Authorization: Bearer $PARALLEL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "parallel",
    "input": "Who is the current CEO of the largest cloud provider by revenue?",
    "reasoning": {"effort": "low"}
  }'

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["PARALLEL_API_KEY"],
    base_url="https://api.parallel.ai/v1",
)

response = client.responses.create(
    model="parallel",
    input="Who is the current CEO of the largest cloud provider by revenue?",
    reasoning={"effort": "low"},  # low | medium | high
)

print(response.output_text)

import OpenAI from "openai";

const client = new OpenAI({
  apiKey: process.env.PARALLEL_API_KEY,
  baseURL: "https://api.parallel.ai/v1",
});

const response = await client.responses.create({
  model: "parallel",
  input: "Who is the current CEO of the largest cloud provider by revenue?",
  reasoning: { effort: "low" },
});

console.log(response.output_text);

Read the answer from output_text (convenience accessor), or walk output[] by item type: one web_search_call item per search the model ran or page it read, then the message item whose content[] carries the answer text.

Migrate from OpenAI

Migrating an existing OpenAI Responses integration is a three-line diff: point base_url at Parallel, swap the key, and set model="parallel". Grounding is automatic, so you can drop the web_search tool. If you keep it, its domain filters are honored (see Web Search Tool):

- client = OpenAI(api_key=OPENAI_API_KEY)
+ client = OpenAI(api_key=PARALLEL_API_KEY, base_url="https://api.parallel.ai/v1")

  response = client.responses.create(
-     model="gpt-5.6-luna",
+     model="parallel",
      input=PROMPT,
      reasoning={"effort": "low"},
-     tools=[{"type": "web_search"}],
  )

See OpenAI Responses Compatibility for the full field-by-field compatibility reference.

Use it as a research subagent

You can use the Responses API as a web research tool for your own agent by creating a tool that wraps the Responses API. Each tool call your agent makes is delegated to an autonomous research subagent that returns a finished, cited answer — not raw search results your agent has to read through.

def web_research(query: str, effort: str = "low") -> dict:
    """Tool handler: one complete question in, one researched answer out."""
    r = parallel.responses.create(
        model="parallel", input=query, reasoning={"effort": effort}
    )
    return {"answer": r.output_text}

Your agent plans and decomposes; Parallel researches. See Using It as a Research Subagent for the tool definition, best practices for writing the tool description, and a full orchestrator loop.

Features

Statefulness

Multi-turn conversations via previous_response_id

Structured Outputs

JSON answers conforming to your schema

Streaming Events

Server-sent events with the standard Responses event sequence

Citations

Source annotations for grounded answers

Web Search Tool

Domain filters and the trail of searches and page reads behind each answer

Examples

Direct Requests

Call it directly: questions, deep research, enrichment, follow-ups

Research Subagent

Expose it as a web-research tool of another LLM

Next steps

Review OpenAI Responses Compatibility for honored, ignored, and rejected request fields
For long-running or batch research, use the Task API — the Responses API is synchronous only
Check pricing and rate limits

Getting started

Web Tools

Web Agents

Developer Tools

Platform Integrations

Data Integrations

Additional Resources

Responses API Quickstart

Model and reasoning tiers

Prerequisites

Make your first request

Migrate from OpenAI

Use it as a research subagent

Features

Statefulness

Structured Outputs

Streaming Events

Citations

Web Search Tool

Examples

Direct Requests

Research Subagent

Next steps

​Model and reasoning tiers

​Prerequisites

​Make your first request

​Migrate from OpenAI

​Use it as a research subagent

​Features

Statefulness

Structured Outputs

Streaming Events

Citations

Web Search Tool

​Examples

Direct Requests

Research Subagent

​Next steps

Model and reasoning tiers

Prerequisites

Make your first request

Migrate from OpenAI

Use it as a research subagent

Features

Examples

Next steps