Databricks Release Notes

Billing logs now record the performance mode of serverless jobs and pipelines. The workload's performance mode is logged in the product_features.performance_target column and can include values of PERFORMANCE_OPTIMIZED, STANDARD, or null.

For billing log reference, see Billable usage system table reference.

• Legend selection for aliased series names: Legend selection now works correctly for charts with aliased series names in SQL editor and notebooks.

The Databricks Assistant can now also use models served through Anthropic on Databricks as part of the Data Science Agent when partner-powered AI features are enabled. Anthropic on Databricks uses endpoints hosted by Databricks within the Databricks security perimeter.

New maintenance updates are available for supported Databricks Runtime versions. These updates include bug fixes, security patches, and performance improvements. For details, see Databricks Runtime maintenance updates.

Databricks Runtime 17.3 LTS and Databricks Runtime 17.3 LTS ML are now in Beta, powered by Apache Spark 4.0.0. The release includes new configuration options, improved error handling, and enhanced Spark Connect support.

See Databricks Runtime 17.3 LTS and Databricks Runtime 17.3 LTS for Machine Learning.

Model Serving now supports the OpenAI GPT-5, GPT-5 mini and GPT-5 nano models in Public Preview. Reach out to your account teams to access these models during the preview.

These models are optimized for AI Functions, which means you can perform batch inference using these models and AI Functions like ai_query().

For real-time inference workloads, see the following pages:

• Query a chat model
• Query reasoning models
• Query vision models
• Function calling

You can now enable partition metadata logging, a partition discovery strategy for external tables registered to Unity Catalog. See Use partition metadata logging.

Delta Sharing recipients can now apply their own row filters and columns masks on shared tables and shared foreign tables. However, Delta Sharing providers still cannot share data assets that have row-level security or column masks.

For details, see Apply row filters and column masks.

You can now apply the system.certification_status governed tag to catalogs, schemas, tables, views, volumes, dashboards, registered models, and Genie Agents to indicate whether a data asset is certified or deprecated. This improves governance, discoverability, and trust in analytics and AI workloads. See Flag data as certified or deprecated.

Prompt caching is now supported for Databricks-hosted Claude models. You can specify the cache_control parameter in your query requests to cache the following:

• Thinking messages content in the messages.content array.
• Images content blocks in the messages.content array.
• Tool use, results and definitions in the tools array.

See Foundation model REST API reference.

Mosaic AI Model Serving now supports Anthropic Claude Sonnet 4.5 as a Databricks-hosted model. You can access this model using Foundation Model APIs pay-per-token.

Mosaic AI Model Serving now supports Anthropic Claude Sonnet 4 for batch inference workflows. You can now use databricks-claude-sonnet-4 in your ai_query requests to perform batch inference.

• See Transform unstructured data using AI Functions for an example.
• See Databricks-hosted inference for region availability.

You can now define semantic metadata in a metric view. Semantic metadata helps AI tools such as Genie Agents (formerly Genie Spaces) and AI/BI dashboards interpret and use your data more effectively.

To use semantic metadata, your metric view must use YAML specification version 1.1 or higher and run on Databricks Runtime 17.3 or above. The corresponding Databricks SQL version is 2025.30, available in the Preview channel for SQL warehouses.

See Agent metadata in metric views and Upgrade considerations.

• Add tables and metric views to spaces: Users can now add tables and metric views to a space and analyze them together.
• Benchmarks scoring improvement: Benchmarks now treat extra rows or columns in Genie's output as incorrect. See Chat mode ratings.
• Thinking steps in responses: Each Genie response now includes thinking steps that show how the prompt was interpreted, along with the tables and example SQL statements that were used. See Review a response.

The ALTER TABLE ... SET MANAGED command is now generally available. This command seamlessly converts Unity Catalog external tables to managed tables. It allows you to take full advantage of Unity Catalog managed table features, such as enhanced governance, reliability, and performance. See Convert external or foreign Delta Lake tables to Unity Catalog managed tables.

You can now define semantic metadata in a metric view. Semantic metadata helps AI tools such as Genie Agents (formerly Genie Spaces) and AI/BI dashboards interpret and use your data more effectively.

To use semantic metadata, your metric view must use YAML specification version 1.1 or higher and run on Databricks Runtime 17.3 or above. The corresponding Databricks SQL version is 2025.30, available in the Preview channel for SQL warehouses.

See Agent metadata in metric views and Upgrade to YAML 1.1.