AI models maintenance policy

This page describes the model maintenance policy for the Foundation Model APIs pay-per-token, Foundation Model APIs provisioned throughput, and Batch Inference with ai_query offerings.

To continue supporting the most state-of-the-art models, Databricks manages models through a lifecycle that progresses from legacy status to deprecation to eventual retirement.

• Legacy: When a model is marked legacy, it is no longer recommended for new workloads. Customers should begin evaluating newer models and consider migration. The model remains available only to workspaces that used it in the 30 days before the legacy announcement.
• Deprecation: When a model is deprecated, Databricks announces a retirement date 30 days or 90 days in the future. The model remains available only to workspaces that used it in the 30 days before the deprecation announcement. Customers using a deprecated model must migrate their workloads before the retirement date to avoid disruption.
• Retirement: When a model is retired, it is no longer accessible and support for the model is fully discontinued. Any workload using the model no longer works.

Model lifecycle policy

The following section summarizes the model lifecycle policy. See Legacy, deprecated, and retired models for a list of legacy and deprecated models and their retirement dates.

Many factors are considered when determining the retirement timeline, including model popularity, suitable replacement model, and amount of production usage. Databricks announces retirement dates for models when they are deprecated, following the notification timelines outlined below.

Foundation Model APIs

The following table summarizes the model lifecycle policy for Foundation Model APIs products: pay-per-token, provisioned throughput, and ai_query (batch inference).

Partner model retirement policy

Partner models are models that third-party partners like OpenAI, Anthropic, and Google, provide through Foundation Model APIs. For these partner models, Databricks generally follows the same deprecation timelines and policies described above.

However, partners might provide retirement dates shorter than the transition period that Databricks publishes. In these cases, Databricks attempts to bridge the gap by temporarily redirecting models to a similar version, so customers receive the full transition time.

For example, if a partner announces a retirement with two weeks' lead time instead of 30 days, Databricks redirects the model for an additional two weeks to prevent immediate breakage and allow time for migration. Queries fail at the end of the full 30-day period.

Legacy, deprecated, and retired models

The following table lists legacy, deprecated, and retired models, with their retirement dates and recommended replacement models. Customers must migrate applications to use replacement models before any indicated retirement date.

Find workloads that use retired models

Use the following query to find workloads that are using deprecated models and identify their owners.

SELECT
   eu.requester,
   se.endpoint_name,
   se.entity_name,
   COUNT(*) AS request_count,
   SUM(eu.input_token_count) AS total_input_tokens,
   SUM(eu.output_token_count) AS total_output_tokens,
   MIN(eu.request_time) AS first_request,
   MAX(eu.request_time) AS last_request
 FROM system.serving.endpoint_usage eu
 JOIN system.serving.served_entities se
   ON eu.served_entity_id = se.served_entity_id
 WHERE LOWER(se.entity_name) LIKE '%<retired-model-name>%'
 GROUP BY eu.requester, se.endpoint_name, se.entity_name
 ORDER BY request_count DESC

If you use legacy foundation and embedding model serving endpoints, please use the following query:

SELECT *
FROM system.ai_gateway.usage
WHERE endpoint_name LIKE '%<retired-model-name>%';