Model deprecation and retirement for Azure AI Foundry Models

Azure AI Foundry Models in the model catalog are continually refreshed with newer and more capable models. As part of this process, model providers might deprecate and retire their older models, and you might need to update your applications to use a newer model. This document communicates information about the model lifecycle and deprecation timelines and explains how you're informed of model lifecycle stages.

Model lifecycle stages

Models in the model catalog belong to one of these stages:

• Preview
• Generally available
• Legacy
• Deprecated
• Retired

Preview

Models labeled Preview are experimental in nature. A model's weights, runtime, and API schema can change while the model is in preview. Models in preview aren't guaranteed to become generally available. Models in preview have a Preview label next to their name in the model catalog.

Generally available

This stage is the default model stage. Models that don't include a lifecycle label next to their name are generally available and suitable for use in production environments. In this stage, model weights and APIs are fixed. However, model containers or runtimes with vulnerabilities might get patched, but patches don't affect model outputs.

Legacy

Models labeled Legacy are intended for deprecation. You should plan to move to a different model, such as a new, improved model that might be available in the same model family. While a model is in the legacy stage, existing deployments of the model continue to work, and you can create new deployments of the model until the deprecation date.

Deprecated

Models labeled Deprecated are no longer available for new deployments. You can't create any new deployments for the model; however, existing deployments continue to work until the retirement date.

Retired

Models labeled Retired are no longer available for use. You can't create new deployments, and attempts to use existing deployments return <return code> errors.

Notifications

• Models are labeled as Legacy and remain in the legacy state for at least 30 days before being moved to the deprecated state. During this notification period, you can create new deployments as you prepare for deprecation and retirement.

• Models are labeled Deprecated and remain in the deprecated state for at least 90 days before being moved to the retired state. During this notification period, you can migrate any existing deployments to newer or replacement models.

• For each subscription that has a model deployed as a standard deployment or deployed in Foundry Models, members of the owner, contributor, reader, monitoring contributor, and monitoring reader roles receive a notification when a model deprecation is announced. The notification contains the dates when the model enters legacy, deprecated, and retired states. The notification might provide information about possible replacement model options, if applicable.

More notification details for Azure OpenAI in Foundry Models

Additionally, for Azure OpenAI models, customers with active Azure OpenAI deployments receive notice for models with upcoming retirement as follows:

• At model launch, we programmatically designate a "not sooner than" retirement date (typically one year out).
• At least 60 days notice before model retirement for Generally Available (GA) models.
• At least 30 days notice before preview model version upgrades.

Retirements are done on a rolling basis, region by region. Notifications are sent from an unmonitored mailbox, azure-noreply@microsoft.com.

Model availability for Azure OpenAI models

• At least one year of model availability for GA models after the release date of a model in at least one region worldwide.
• For global deployments, all future model versions starting with gpt-4o and gpt-4 0409 will be available with their (N) next succeeding model (N+1) for comparison together.
• Customers have 60 days to try out a new GA model in at least one global, or standard region, before any upgrades happen to a newer GA model.

Considerations for the Azure public cloud

Be aware of the following details for the Azure public cloud :

• All model version combinations will not be available in all regions.
• Model version N and N+1 might not always be available in the same region.
• GA model version N might upgrade to a future model version N+X in some regions based on capacity limitations, and without the new model version N+X separately being available to test in the same region. The new model version will be available to test in other regions before any upgrades are scheduled.
• Preview model versions and GA versions of the same model won't always be available to test together in the same region. There will be preview and GA versions available to test in different regions.
• To balance service quality for existing customers, we reserve the right to limit future customers from using a particular region.
• As always at Microsoft, security is of the utmost importance. If a model or model version is found to have compliance or security issues, we reserve the right to invoke the need to do emergency retirements. For more information, see the terms of service.

Special considerations for Azure Government clouds

• Global standard deployments won't be available in government clouds.
• Not all models or model versions available in commercial/public cloud will be available in government clouds.
• In the Azure Government clouds, we intend to support only one version of a given model at a time.
  • For example, only one version of gpt-35-turbo 0125 and gpt-4o (2024-05-13).
• There will, however, be a 30-day overlap between new model versions, where more than two will be available.
  • For example, if gpt-35-turbo 0125 or gpt-4o (2024-05-13) is updated to a future version, or
  • For model family changes beyond version updates, such as when moving from gpt-4 1106-preview to gpt-4o (2024-05-13).

How to get ready for model retirements and version upgrades

To prepare for model retirements and version upgrades, we recommend that customers test their applications with the new models and versions and evaluate their behavior. We also recommend that customers update their applications to use the new models and versions before the retirement date.

For more information on the model evaluation process, see the Getting started with model evaluation guide.

For information on the model upgrade process, see Model versions in Azure AI Foundry Models and How to upgrade to a new model or version.

For more information on how to manage model upgrades and migrations for provisioned deployments, see Managing models on provisioned deployment types

Timelines for Azure OpenAI models

The following sections list the current, default, and deprecated Azure OpenAI models.

Current Azure OpenAI models

These models are currently available for use in Azure OpenAI.

We'll notify all customers with these preview deployments at least 30 days before the start of the upgrades. We'll publish an upgrade schedule detailing the order of regions and model versions that we'll follow during the upgrades, and link to that schedule from here.

Default Azure OpenAI model versions

Deprecated Azure OpenAI models

These models are no longer available for new deployments.

If you're an existing customer looking for information about these models, see Legacy models.

To track individual updates to this article for Azure OpenAI model retirements, refer to the Git History.

Timelines for other Foundry Models

The following tables list the timelines for models that are on track for retirement. The specified dates are in UTC time.

AI21 Labs

Cohere

DeepSeek

Meta

Mistral AI

Related content

• Explore Azure AI Foundry Models
• Data, privacy, and security for use of models through the model catalog in Azure AI Foundry portal