Azure OpenAI Service models

Azure OpenAI Service is powered by a diverse set of models with different capabilities and price points. Model availability varies by region and cloud. For Azure Government model availability, please refer to Azure Government OpenAI Service.

o1-preview and o1-mini models limited access

The Azure OpenAI o1-preview and o1-mini models are specifically designed to tackle reasoning and problem-solving tasks with increased focus and capability. These models spend more time processing and understanding the user's request, making them exceptionally strong in areas like science, coding, and math compared to previous iterations.

Availability

The o1-preview and o1-mini models are now available for API access and model deployment. Registration is required, and access will be granted based on Microsoft's eligibility criteria.

Request access: limited access model application

Once access has been granted, you will need to create a deployment for each model.

API support

Support for the o1 series models was added in API version 2024-09-01-preview.

The max_tokens parameter has been deprecated and replaced with the new max_completion_tokens parameter. o1 series models will only work with the max_completion_tokens parameter.

Usage

These models do not currently support the same set of parameters as other models that use the chat completions API. Only a very limited subset is currently supported, so common parameters like temperature, top_p, are not available and including them will cause your request to fail. o1-preview and o1-mini models will also not accept the system message role as part of the messages array.

pip install openai --upgrade

from openai import AzureOpenAI
from azure.identity import DefaultAzureCredential, get_bearer_token_provider

token_provider = get_bearer_token_provider(
    DefaultAzureCredential(), "https://cognitiveservices.azure.com/.default"
)

client = AzureOpenAI(
  azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT"), 
  azure_ad_token_provider=token_provider,
  api_version="2024-09-01-preview"
)

response = client.chat.completions.create(
    model="o1-preview-new", # replace with the model deployment name of your o1-preview, or o1-mini model
    messages=[
        {"role": "user", "content": "What steps should I think about when writing my first Python API?"},
    ],
    max_completion_tokens = 5000

)

print(response.model_dump_json(indent=2))

pip install openai --upgrade

from openai import AzureOpenAI

client = AzureOpenAI(
  azure_endpoint = os.getenv("AZURE_OPENAI_ENDPOINT"), 
  api_key=os.getenv("AZURE_OPENAI_API_KEY"),  
  api_version="2024-09-01-preview"
)

response = client.chat.completions.create(
    model="o1-preview-new", # replace with the model deployment name of your o1-preview, or o1-mini model
    messages=[
        {"role": "user", "content": "What steps should I think about when writing my first Python API?"},
    ],
    max_completion_tokens = 5000

)

print(response.model_dump_json(indent=2))

{
  "id": "chatcmpl-AEj7pKFoiTqDPHuxOcirA9KIvf3yz",
  "choices": [
    {
      "finish_reason": "stop",
      "index": 0,
      "logprobs": null,
      "message": {
        "content": "Writing your first Python API is an exciting step in developing software that can communicate with other applications. An API (Application Programming Interface) allows different software systems to interact with each other, enabling data exchange and functionality sharing. Here are the steps you should consider when creating your first Python API:\n\n1. **Define the Purpose and Requirements**\n\n   - **Identify the Functionality**: Clearly outline what your API is supposed to do. What data or services will it provide to the users?\n   - **Determine the Endpoints**: Plan the different URLs (endpoints) through which users can access the API functionalities.\n   - **Specify Request and Response Formats**: Decide on the data formats (usually JSON) for incoming requests and outgoing responses.\n\n2. **Choose the Right Framework**\n\n   Python offers several frameworks for building APIs. Two of the most popular are:\n\n   - **Flask**: A lightweight and flexible web framework, great for small to medium-sized APIs.\n   - **FastAPI**: A modern, high-performance framework for building APIs with Python 3.6+ types, offering automatic interactive documentation.\n\n   **Example**:\n   ```bash\n   pip install flask\n   ```\n   or\n   ```bash\n   pip install fastapi uvicorn\n   ```\n\n3. **Set Up the Development Environment**\n\n   - **Create a Virtual Environment**: Isolate your project dependencies using `venv` or `conda`.\n   - **Install Required Packages**: Ensure all necessary libraries and packages are installed.\n\n   **Example**:\n   ```bash\n   python -m venv env\n   source env/bin/activate  # On Windows use `env\\Scripts\\activate`\n   ```\n\n4. **Implement the API Endpoints**\n\n   - **Write the Code for Each Endpoint**: Implement the logic that handles requests and returns responses.\n   - **Use Decorators to Define Routes**: In frameworks like Flask, you use decorators to specify the URL endpoints.\n\n   **Example with Flask**:\n   ```python\n   from flask import Flask, request, jsonify\n\n   app = Flask(__name__)\n\n   @app.route('/hello', methods=['GET'])\n   def hello_world():\n       return jsonify({'message': 'Hello, World!'})\n\n   if __name__ == '__main__':\n       app.run(debug=True)\n   ```\n\n5. **Handle Data Serialization and Deserialization**\n\n   - **Parsing Incoming Data**: Use libraries to parse JSON or other data formats from requests.\n   - **Formatting Output Data**: Ensure that responses are properly formatted in JSON or XML.\n\n6. **Implement Error Handling**\n\n   - **Handle Exceptions Gracefully**: Provide meaningful error messages and HTTP status codes.\n   - **Validate Input Data**: Check for required fields and appropriate data types to prevent errors.\n\n   **Example**:\n   ```python\n   @app.errorhandler(404)\n   def resource_not_found(e):\n       return jsonify(error=str(e)), 404\n   ```\n\n7. **Add Authentication and Authorization (If Necessary)**\n\n   - **Secure Endpoints**: If your API requires, implement security measures such as API keys, tokens (JWT), or OAuth.\n   - **Manage User Sessions**: Handle user login states and permissions appropriately.\n\n8. **Document Your API**\n\n   - **Use Tools Like Swagger/OpenAPI**: Automatically generate interactive API documentation.\n   - **Provide Usage Examples**: Help users understand how to interact with your API.\n\n   **Example with FastAPI**:\n   FastAPI automatically generates docs at `/docs` using Swagger UI.\n\n9. **Test Your API**\n\n   - **Write Unit and Integration Tests**: Ensure each endpoint works as expected.\n   - **Use Testing Tools**: Utilize tools like `unittest`, `pytest`, or API testing platforms like Postman.\n\n   **Example**:\n   ```python\n   import unittest\n   class TestAPI(unittest.TestCase):\n       def test_hello_world(self):\n           response = app.test_client().get('/hello')\n           self.assertEqual(response.status_code, 200)\n   ```\n\n10. **Optimize Performance**\n\n    - **Improve Response Times**: Optimize your code and consider using asynchronous programming if necessary.\n    - **Manage Resource Utilization**: Ensure your API can handle the expected load.\n\n11. **Deploy Your API**\n\n    - **Choose a Hosting Platform**: Options include AWS, Heroku, DigitalOcean, etc.\n    - **Configure the Server**: Set up the environment to run your API in a production setting.\n    - **Use a Production Server**: Instead of the development server, use WSGI servers like Gunicorn or Uvicorn.\n\n    **Example**:\n    ```bash\n    uvicorn main:app --host 0.0.0.0 --port 80\n    ```\n\n12. **Monitor and Maintain**\n\n    - **Logging**: Implement logging to track events and errors.\n    - **Monitoring**: Use monitoring tools to track performance and uptime.\n    - **Update and Patch**: Keep dependencies up to date and patch any security vulnerabilities.\n\n13. **Consider Versioning**\n\n    - **Plan for Updates**: Use versioning in your API endpoints to manage changes without breaking existing clients.\n    - **Example**:\n      ```python\n      @app.route('/v1/hello', methods=['GET'])\n      ```\n\n14. **Gather Feedback and Iterate**\n\n    - **User Feedback**: Encourage users to provide feedback on your API.\n    - **Continuous Improvement**: Use the feedback to make improvements and add features.\n\n**Additional Tips**:\n\n- **Keep It Simple**: Start with a minimal viable API and expand functionality over time.\n- **Follow RESTful Principles**: Design your API according to REST standards to make it intuitive and standard-compliant.\n- **Security Best Practices**: Always sanitize inputs and protect against common vulnerabilities like SQL injection and cross-site scripting (XSS).\nBy following these steps, you'll be well on your way to creating a functional and robust Python API. Good luck with your development!",
        "refusal": null,
        "role": "assistant",
        "function_call": null,
        "tool_calls": null
      },
      "content_filter_results": {
        "hate": {
          "filtered": false,
          "severity": "safe"
        },
        "protected_material_code": {
          "filtered": false,
          "detected": false
        },
        "protected_material_text": {
          "filtered": false,
          "detected": false
        },
        "self_harm": {
          "filtered": false,
          "severity": "safe"
        },
        "sexual": {
          "filtered": false,
          "severity": "safe"
        },
        "violence": {
          "filtered": false,
          "severity": "safe"
        }
      }
    }
  ],
  "created": 1728073417,
  "model": "o1-preview-2024-09-12",
  "object": "chat.completion",
  "service_tier": null,
  "system_fingerprint": "fp_503a95a7d8",
  "usage": {
    "completion_tokens": 1843,
    "prompt_tokens": 20,
    "total_tokens": 1863,
    "completion_tokens_details": {
      "audio_tokens": null,
      "reasoning_tokens": 448
    },
    "prompt_tokens_details": {
      "audio_tokens": null,
      "cached_tokens": 0
    }
  },
  "prompt_filter_results": [
    {
      "prompt_index": 0,
      "content_filter_results": {
        "custom_blocklists": {
          "filtered": false
        },
        "hate": {
          "filtered": false,
          "severity": "safe"
        },
        "jailbreak": {
          "filtered": false,
          "detected": false
        },
        "self_harm": {
          "filtered": false,
          "severity": "safe"
        },
        "sexual": {
          "filtered": false,
          "severity": "safe"
        },
        "violence": {
          "filtered": false,
          "severity": "safe"
        }
      }
    }
  ]
}

Region availability

Available for standard and global standard deployment in East US, East US2, North Central US, South Central US, Sweden Central, West US, and West US3 for approved customers.

GPT-4o-Realtime-Preview

The gpt-4o-realtime-preview model is part of the GPT-4o model family and supports low-latency, "speech in, speech out" conversational interactions. GPT-4o audio is designed to handle real-time, low-latency conversational interactions, making it a great fit for support agents, assistants, translators, and other use cases that need highly responsive back-and-forth with a user.

GPT-4o audio is available in the East US 2 (eastus2) and Sweden Central (swedencentral) regions. To use GPT-4o audio, you need to create or use an existing resource in one of the supported regions.

When your resource is created, you can deploy the GPT-4o audio model. If you are performing a programmatic deployment, the model name is gpt-4o-realtime-preview. For more information on how to use GPT-4o audio, see the GPT-4o audio documentation.

Details about maximum request tokens and training data are available in the following table.

GPT-4o and GPT-4 Turbo

GPT-4o integrates text and images in a single model, enabling it to handle multiple data types simultaneously. This multimodal approach enhances accuracy and responsiveness in human-computer interactions. GPT-4o matches GPT-4 Turbo in English text and coding tasks while offering superior performance in non-English languages and vision tasks, setting new benchmarks for AI capabilities.

How do I access the GPT-4o and GPT-4o mini models?

GPT-4o and GPT-4o mini are available for standard and global-standard model deployment.

You need to create or use an existing resource in a supported standard or global standard region where the model is available.

When your resource is created, you can deploy the GPT-4o models. If you are performing a programmatic deployment, the model names are:

• gpt-4o Version 2024-08-06
• gpt-4o, Version 2024-05-13
• gpt-4o-mini Version 2024-07-18

GPT-4 Turbo

GPT-4 Turbo is a large multimodal model (accepting text or image inputs and generating text) that can solve difficult problems with greater accuracy than any of OpenAI's previous models. Like GPT-3.5 Turbo, and older GPT-4 models GPT-4 Turbo is optimized for chat and works well for traditional completions tasks.

The latest GA release of GPT-4 Turbo is:

• gpt-4 Version: turbo-2024-04-09

This is the replacement for the following preview models:

• gpt-4 Version: 1106-Preview
• gpt-4 Version: 0125-Preview
• gpt-4 Version: vision-preview

Differences between OpenAI and Azure OpenAI GPT-4 Turbo GA Models

• OpenAI's version of the latest 0409 turbo model supports JSON mode and function calling for all inference requests.
• Azure OpenAI's version of the latest turbo-2024-04-09 currently doesn't support the use of JSON mode and function calling when making inference requests with image (vision) input. Text based input requests (requests without image_url and inline images) do support JSON mode and function calling.

Differences from gpt-4 vision-preview

• Azure AI specific Vision enhancements integration with GPT-4 Turbo with Vision isn't supported for gpt-4 Version: turbo-2024-04-09. This includes Optical Character Recognition (OCR), object grounding, video prompts, and improved handling of your data with images.

GPT-4 Turbo provisioned managed availability

• gpt-4 Version: turbo-2024-04-09 is available for both standard and provisioned deployments. Currently the provisioned version of this model doesn't support image/vision inference requests. Provisioned deployments of this model only accept text input. Standard model deployments accept both text and image/vision inference requests.

Deploying GPT-4 Turbo with Vision GA

To deploy the GA model from the Studio UI, select GPT-4 and then choose the turbo-2024-04-09 version from the dropdown menu. The default quota for the gpt-4-turbo-2024-04-09 model will be the same as current quota for GPT-4-Turbo. See the regional quota limits.

GPT-4

GPT-4 is the predecessor to GPT-4 Turbo. Both the GPT-4 and GPT-4 Turbo models have a base model name of gpt-4. You can distinguish between the GPT-4 and Turbo models by examining the model version.

• gpt-4 Version 0314
• gpt-4 Version 0613
• gpt-4-32k Version 0613

You can see the token context length supported by each model in the model summary table.

GPT-4 and GPT-4 Turbo models

• These models can only be used with the Chat Completion API.

See model versions to learn about how Azure OpenAI Service handles model version upgrades, and working with models to learn how to view and configure the model version settings of your GPT-4 deployments.

• GPT-4 version 0125-preview is an updated version of the GPT-4 Turbo preview previously released as version 1106-preview.
• GPT-4 version 0125-preview completes tasks such as code generation more completely compared to gpt-4-1106-preview. Because of this, depending on the task, customers may find that GPT-4-0125-preview generates more output compared to the gpt-4-1106-preview. We recommend customers compare the outputs of the new model. GPT-4-0125-preview also addresses bugs in gpt-4-1106-preview with UTF-8 handling for non-English languages.
• GPT-4 version turbo-2024-04-09 is the latest GA release and replaces 0125-Preview, 1106-preview, and vision-preview.

GPT-3.5

GPT-3.5 models can understand and generate natural language or code. The most capable and cost effective model in the GPT-3.5 family is GPT-3.5 Turbo, which has been optimized for chat and works well for traditional completions tasks as well. GPT-3.5 Turbo is available for use with the Chat Completions API. GPT-3.5 Turbo Instruct has similar capabilities to text-davinci-003 using the Completions API instead of the Chat Completions API. We recommend using GPT-3.5 Turbo and GPT-3.5 Turbo Instruct over legacy GPT-3.5 and GPT-3 models.

To learn more about how to interact with GPT-3.5 Turbo and the Chat Completions API check out our in-depth how-to.

¹ This model will accept requests > 4,096 tokens. It is not recommended to exceed the 4,096 input token limit as the newer version of the model are capped at 4,096 tokens. If you encounter issues when exceeding 4,096 input tokens with this model this configuration is not officially supported.

Embeddings

text-embedding-3-large is the latest and most capable embedding model. Upgrading between embeddings models is not possible. In order to move from using text-embedding-ada-002 to text-embedding-3-large you would need to generate new embeddings.

• text-embedding-3-large
• text-embedding-3-small
• text-embedding-ada-002

In testing, OpenAI reports both the large and small third generation embeddings models offer better average multi-language retrieval performance with the MIRACL benchmark while still maintaining performance for English tasks with the MTEB benchmark.

The third generation embeddings models support reducing the size of the embedding via a new dimensions parameter. Typically larger embeddings are more expensive from a compute, memory, and storage perspective. Being able to adjust the number of dimensions allows more control over overall cost and performance. The dimensions parameter is not supported in all versions of the OpenAI 1.x Python library, to take advantage of this parameter we recommend upgrading to the latest version: pip install openai --upgrade.

OpenAI's MTEB benchmark testing found that even when the third generation model's dimensions are reduced to less than text-embeddings-ada-002 1,536 dimensions performance remains slightly better.

DALL-E

The DALL-E models generate images from text prompts that the user provides. DALL-E 3 is generally available for use with the REST APIs. DALL-E 2 and DALL-E 3 with client SDKs are in preview.

Whisper

The Whisper models can be used for speech to text.

You can also use the Whisper model via Azure AI Speech batch transcription API. Check out What is the Whisper model? to learn more about when to use Azure AI Speech vs. Azure OpenAI Service.

Text to speech (Preview)

The OpenAI text to speech models, currently in preview, can be used to synthesize text to speech.

You can also use the OpenAI text to speech voices via Azure AI Speech. To learn more, see OpenAI text to speech voices via Azure OpenAI Service or via Azure AI Speech guide.

Model summary table and region availability

Models by deployment type

Azure OpenAI provides customers with choices on the hosting structure that fits their business and usage patterns. The service offers two main types of deployment:

• Standard is offered with a global deployment option, routing traffic globally to provide higher throughput.
• Provisioned is also offered with a global deployment option, allowing customers to purchase and deploy provisioned throughput units across Azure global infrastructure.

All deployments can perform the exact same inference operations, however the billing, scale, and performance are substantially different. To learn more about Azure OpenAI deployment types see our deployment types guide.

This table doesn't include fine-tuning regional availability information. Consult the fine-tuning section for this information.

Standard models by endpoint

Fine-tuning models

¹ GPT-4 is currently in public preview.

Assistants (Preview)

For Assistants you need a combination of a supported model, and a supported region. Certain tools and capabilities require the latest models. The following models are available in the Assistants API, SDK, and Azure AI Studio. The following table is for pay-as-you-go. For information on Provisioned Throughput Unit (PTU) availability, see provisioned throughput. The listed models and regions can be used with both Assistants v1 and v2. You can use global standard models if they are supported in the regions listed below.

Model retirement

For the latest information on model retirements, refer to the model retirement guide.

Next steps

• Model retirement and deprecation
• Learn more about working with Azure OpenAI models
• Learn more about Azure OpenAI
• Learn more about fine-tuning Azure OpenAI models