Grok API with Python

Updated: 2026-08-15

A working Grok 4.5 integration for Python using the OpenAI SDK and the Grok API Dev base URL. The examples cover a regular request, AsyncOpenAI, streaming, timeouts, and API errors.

Install the Python SDK

Use the current OpenAI package in a virtual environment. Keep the API key outside the source code and pin the tested dependency version in your requirements file or lockfile.

Bash

python -m pip install --upgrade openai

Bash

export GROK_API_DEV_KEY="sk-lg-YOUR_API_KEY"

The base_url already includes /v1. Do not append /responses when creating the client because the SDK adds the method path itself.

Send the first request

The Responses API accepts a string in input for a minimal text request. The SDK exposes the collected answer through output_text and token accounting through usage.

Python

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["GROK_API_DEV_KEY"],
    base_url="https://api.llm-gate.tech/v1",
    timeout=3600.0,
    max_retries=2,
)

response = client.responses.create(
    model="grok-4.5",
    input="Explain SSE in two sentences.",
)

print(response.output_text)
print(response.usage)

Run this code on a server, in a local script, or in a trusted worker. The API key must not be shipped to browser code.

Use AsyncOpenAI

The asynchronous client fits FastAPI, background workers, and services that already use asyncio. Awaiting one request does not limit concurrency on its own, so apply a queue or semaphore when traffic can spike.

Python async

import asyncio
import os
from openai import AsyncOpenAI

client = AsyncOpenAI(
    api_key=os.environ["GROK_API_DEV_KEY"],
    base_url="https://api.llm-gate.tech/v1",
)

async def main() -> None:
    response = await client.responses.create(
        model="grok-4.5",
        input="Hello!",
    )
    print(response.output_text)

asyncio.run(main())

Reuse one client instead of creating a new HTTP connection for every request.

Stream a Grok response

Set stream=True to receive events while the answer is being generated. Append only response.output_text.delta events to the visible text and wait for response.completed before marking the request as finished.

Python streaming

stream = client.responses.create(
    model="grok-4.5",
    input="Write a short Python example.",
    stream=True,
)

for event in stream:
    if event.type == "response.output_text.delta":
        print(event.delta, end="", flush=True)

A stream can return an error after the connection has opened. Keep the partial text separate from a completed response.

Timeouts, retries, and errors

The OpenAI SDK retries selected temporary failures. Set a finite timeout and max_retries so request behavior stays predictable.