Databricks

Databricks and OpenAI announced a partnership that makes OpenAI's GPT-5 model natively available on the Databricks Data + AI Platform, as confirmed by the official Databricks blog post. Customers can call GPT-5 from within SQL, as a Model Serving API endpoint, or through Agent Bricks, without additional setup. The integration enables governed, secure access to GPT-5 alongside enterprise data on Databricks. GPT-5 delivers major advances in instruction following and agentic tool use, allowing it to reliably execute multi-step requests and coordinate across tools, according to the Databricks blog. Agent Bricks adds an evaluation loop that measures GPT-5 accuracy on domain-specific tasks and tunes it for higher quality. Documented use cases include healthcare reasoning agents, productivity agents, and coding agents that leverage GPT-5's extended reasoning mode.

Databricks published its Foundation Model APIs model maintenance policy (updated Sep 17, 2026), describing the lifecycle governing pay-per-token, provisioned throughput, Batch Inference with ai query, and fine-tuning offerings. Models progress through update, deprecated, and retired phases, with deprecation signaling t The policy outlines customer notification steps: when a model is deprecated, the Databricks UI surfaces a warning message and the documentation carries a deprecation notice with any announced retirement date. It also points to separate "Deprecated and retired models" and "Partner model retirement policy" pages for conc

Databricks shipped a wave of AI/BI dashboard and Genie One improvements on September 17, 2026, focusing on dashboard authoring and visualization flexibility. Developers can now convert a SQL dataset to a local metric view via the dataset's context menu, with Genie Code handling the conversion, and Genie Code for dashbo Genie One skills gained support for multiple files and sub-folders (up to three folders deep and 50 files or folders per skill), limited to text files, which broadens how teams package knowledge for Genie agents. Other shipped fixes and features include styling of pivot table totals (headers, values, and conditional st

Databricks posted its "What's coming" roadmap (last updated Sep 17, 2026) outlining upcoming behavioral changes for the platform on AWS. Notable items include the removal of Anthropic models from AWS GovCloud IL5 and GovCloud DoD workspaces on September 19, 2026, after which IL5 workspaces will be unable to create or u Additional roadmap items include auto-enablement of JAR tasks on serverless compute for compliance-profile workspaces in early October 2026, the rollout of automatic change data feed (Auto CDF) to all supported regions by the end of October 2026 (requiring Databricks Runtime 19+ with row tracking), and upcoming horizon

Databricks Foundation Model APIs documentation, last updated September 23, 2026, catalogs the open and third-party models available through its pay-per-token endpoints and AI Playground. The page lists OpenAI, Google Gemini, and Anthropic models, with context windows and maximum output tokens matching the values published by each model provider. For production workloads, Databricks recommends provisioned throughput mode. Two OpenAI GPT-6 family endpoints, databricks-gpt-6-sol and databricks-gpt-6-luna, are exposed as pay-per-token endpoints supporting text and image inputs. GPT-6 Sol is described as OpenAI's latest general-purpose model for agentic coding, long-horizon reasoning, and complex tool use, while GPT-6 Luna is positioned as the cost-efficient variant with strong reasoning. Both require the Responses API and are not available in AI Playground.

The Microsoft Learn mirror of Azure Databricks' AI/BI and Genie One release notes documents September 10, 2026 updates to Genie Agents, including raising the data-source limit per agent from 30 to 50 tables, views, or metric views, and increasing the conversation storage limit per agent from 10,000 to 200,000 conversat Dashboard enhancements on the same date include support for more parameterized Tableau calculations via the /importBI Genie Code command, relationship-scoped filters (Public Preview) that restrict widget responses to related datasets, and the ability to choose which SQL warehouse materializes a dashboard dataset export