The OpenAI SDK for Grok

Updated November 3, 2026, based on the xAI Models page: grok-4.7 (September 2026) succeeded grok-4.5 as the recommended model; the example, written with grok-4.5, also works with grok-4.7. Sources: Models · Release notes.

Clarified November 3, 2026, based on the xAI Chat Completions reference: the call isn’t strictly identical to OpenAI’s; logprobs and top_logprobs (ignored from grok-4.20 on) and logit_bias (not supported) add to the differences listed below. Source: xAI, Chat Completions.

Reuse what you know about OpenAI

If you already work with the OpenAI SDK in Python, you can use it as is to call the Grok API. Just redirect requests to xAI’s servers by changing the base URL. It’s the fastest approach if you have existing code.

Installation

If the OpenAI SDK isn’t installed in your project yet:

pip install openai

Configure the client for xAI

The only difference from regular OpenAI usage is the base_url parameter:

from openai import OpenAI
import os

client = OpenAI(
    api_key=os.getenv("XAI_API_KEY"),
    base_url="https://api.x.ai/v1"
)

That’s it. From here on, every method in the OpenAI SDK works with the Grok models.

Send a message

The call has the same shape as for GPT, except for unsupported parameters (see below):

response = client.chat.completions.create(
    model="grok-4.5",
    messages=[
        {"role": "system", "content": "You are a technical assistant."},
        {"role": "user", "content": "What's the difference between REST and GraphQL?"}
    ],
    temperature=0.7,
    max_completion_tokens=500
)

print(response.choices[0].message.content)

Available models

Just replace the OpenAI model name with a Grok model:

• grok-4.7: the model xAI recommends, versatile, with image input (grok-4.6 and grok-4.5 remain available)
• grok-4.20-0309-reasoning: advanced reasoning
• grok-4.3: fast, inexpensive answers (1M context)

Migrate an existing OpenAI project

To adapt an existing project, you have two options.

Option 1: Environment variables

Change your variables without touching the code:

export OPENAI_API_KEY="xai-your-key"
export OPENAI_BASE_URL="https://api.x.ai/v1"

The OpenAI SDK reads these variables automatically. Your existing code will work if you also change the model name and remove the parameters xAI doesn’t support.

Option 2: Explicit configuration

Write an initialization function that switches between providers:

import os
from openai import OpenAI

def create_client(provider="xai"):
    if provider == "xai":
        return OpenAI(
            api_key=os.getenv("XAI_API_KEY"),
            base_url="https://api.x.ai/v1"
        )
    return OpenAI()  # OpenAI by default

client = create_client("xai")

This approach lets you switch easily between providers during the migration.

Differences to know

Although the API is compatible, some behavior differs:

• The frequency_penalty and presence_penalty parameters are not supported on reasoning models
• The stop parameter isn’t available on reasoning models
• logprobs and top_logprobs are silently ignored by grok-4.20 and newer models, and logit_bias isn’t supported
• Model names are obviously different
• Some xAI features (X search, web search) aren’t available through the standard OpenAI format

When to choose this approach

Compatibility mode makes the most sense when your existing code matters more than exclusive features. The most common case is migration: code already written for OpenAI switches to Grok by changing the base URL and the key: two lines, no rewrite. Reversibility matters too: as long as you stay on the OpenAI format, switching providers remains a matter of configuration, which is valuable for comparing models or negotiating. Then there’s the ecosystem: third-party tools built around the OpenAI SDK, such as LangChain and LlamaIndex, work immediately, and a team that already knows the SDK has nothing new to learn. The price to pay is the one you saw above: xAI-specific features take a back seat.

Key takeaways

• Add base_url="https://api.x.ai/v1" to the OpenAI() constructor
• Replace the API key and the model name
• The OPENAI_BASE_URL and OPENAI_API_KEY variables allow a migration without touching the code
• Some parameters (penalties, stop) aren’t supported on reasoning models