Choose the primary AI model for the agent.

Using Copilot Studio , you can select the best model for your agent dispatching using a simple drop-down menu.

Want to test advanced prototypes before they're ready for production? Access the latest prototypes to evaluate them early. However, these may have limitations in terms of testing, availability, and functionality.

This article describes how to select an AI model for agent generation orchestration. There are separate settings to change the model for deep inference (preview), generative feedback (preview), and prompt generator.

Important note :

1. Experimental models are available for exploration and testing but are not recommended for use in a production environment. Review the limitations of experimental models and previews before selecting an experimental model or preview for your agent.
2. Data processed in the test model may be processed and stored outside the geographic location of your organization.
3. This article contains Copilot Studio documentation on model selection, including previews of experimental models, and is subject to change.

Model availability by region

Copilot Studio offers a variety of different models. These models are categorized based on their intended use and availability.

You can view the tags for each model in the model list in Copilot Studio.

The following tables show the availability status of selected models across specific regions and areas.

Public availability

Note : Models marked as geographicly cross-regional may process data outside your region.

Important note :

Microsoft's responsible AI and safety reviews found Grok-4.1 Fast (Non-Reasoning) to be less suitable than other models evaluated, resulting in (i) a higher risk of the model generating potentially harmful content and (ii) lower scores in device safety and jailbreak standards.

Grok-4.1 Fast (Non-Reasoning) can generate pornographic content, and can do so more frequently than other models. Customers must comply with both the Microsoft Enterprise AI Service Code of Conduct and the xAI Enterprise Terms of Service, including the Acceptable Use Policy.

Additionally, there may be types of harm that this model could cause that Microsoft's content safety system does not cover. Therefore, like all Trial models, Grok-4.1 Fast (Non-Reasoning) is not recommended for use in production environments, and customers should consider the limitations of trial and preview models, and conduct their own evaluation before choosing Grok-4.1 Fast (Non-Reasoning).

Classify the purpose of using the model.

Models are optimized for different purposes. Your agent may perform better when you choose a model whose strengths align with the agent's purpose. For example, an agent making complex decisions might benefit from a Deep model, while an agent expected to discuss a wide range of topics might use a General model.

The following table describes the tags that use the model, their strengths, and things to keep in mind if you use that model.

Model versions

Each model listed in Copilot Studio has a tag that identifies its version type. You can try advanced, modern test and preview models, or choose a reliable, thoroughly tested model that is available for official use.

1. Testing : For testing purposes only and not intended for use in a production environment. Subject to preview terms and may have limitations on availability and quality. 
2. Preview : This may become a model available for official use in the future, but is not currently intended for production use. Subject to preview terms and may have limitations on availability and quality. 
3. Available for Official Use (GA) : Models without a version tag are available for official use. You can use this model for large-scale and production applications. In most cases, models available for official use have no restrictions on availability and quality, but some may still have some limitations, such as regional availability.
4. Default : The default model for all agents and is usually the best-performing model among those available for official use. The default model is upgraded periodically as newer, more capable models become available for official use. Agents also use the default model as a fallback option if the chosen model is disabled or unavailable.
5. Deactivated : When a new model becomes the default model, the old default model is deactivated. You can still use the deactivated model for up to one month after it is deactivated. 
6. Geographic regions : This may require processing and storing data outside your organization's geographic boundaries. Your administrators can enable or disable data movement between regions.
7. Early access environments : Environments with an early release cycle. These environments receive updates and new features first. 

External models

You can also add external AI models from Anthropic or xAI to your agent.

Limitations of experimental and preview models

You can explore and experiment with prototype and preview models, but do not use them in a production environment:

1. These may indicate performance fluctuations, response quality issues, latency, or message consumption, and may show timeouts or unavailability.
2. If you publish an agent with a test or preview model, and users use that agent, that usage will be charged at the established price.

Feel free to experiment with these models. However, exercise caution when deploying them in a production environment.

Test and preview models are subject to review terms. Microsoft provides these models before the official release so you can access them early and provide feedback. 

Change the agent's AI model.

Your agent starts with a default model optimized for most cases. To change the agent's model:

1. Access the agent's Overview page .
2. In the Model section , select the primary model of the agent. You can switch between test and production models at any time.

Administrator control over AI model selection

Administrators can allow or block users from adding AI preview and testing models to the agent using the following settings:

1. Administrators can choose to allow or block preview and experimental models in an environment. To use these models, the Preview and experimental AI models setting must be enabled for your environment.
2. Data processed in preview or test models can be processed and stored outside your organization's geographic location. To use test models, your environment must have the " Move data across regions" setting enabled . Tenant administrators manage this environment-level setting in the Power Platform admin center.

Control and governance requirements for external models

Administrators control whether creators can add external models to the agent. To grant access to external models, administrators must perform the following actions:

1. Enable external models in the Power Platform admin center for an environment or group of environments.
2. Allows access to each external model vendor within the Microsoft 365 admin center.

Preview models and external models are two different sets that may overlap but are not identical, and their settings are separate. For example:

1. Administrators can block external models but allow preview or experimental models. In this case, creators cannot use external models but can use available preview, experimental, and internal models.
2. Administrators can block preview or test models but allow external models. In this case, creators cannot use any preview or test models, but can use any available external and internal models.