Unity Gateway became generally available on 4 August 2026. Databricks describes it as the governance solution for enterprise AI, part of Unity Catalog, and the release note lists four things it does. Teams can control which AI services are available to them, route and manage AI traffic across providers, govern MCP servers to control access and costs, and monitor usage, cost, access and lineage from one place. The same release note adds a sentence that decides how you should plan around it, because some capabilities, including service policies and agent services, remain in Beta.
That split is the whole story. GA covers the control plane itself, so the gateway is a supported place to put your model traffic and your MCP servers. The controls that most teams describe when they say they want an AI gateway are a mix of GA and Beta, and the mix keeps moving.
Which capabilities carry which stage #
Budgets went generally available on 6 July 2026, including budgets for Unity Gateway, with per-user thresholds and overrides for requests managed through the gateway and configurable actions including email alerting and usage blocking. That is the cost control, and it's supported. The Unity Gateway API for managing endpoints, model providers and MCP servers arrived on 6 August 2026 across the Python, Go and Java SDKs, the Databricks CLI and the Terraform provider, so the gateway can be configured from code rather than the UI.
Service policies and agent services sit on the other side of the line, named in the GA release note as still Beta. Databricks-managed MCP connectors for Genie One and Genie Code migrated into the gateway on 6 August 2026, also in Beta. The unified trace table, the one place that captures every request and response across gateway services, is in Beta too, and its beta capabilities are enabled separately from an account admin's Previews page. A generally available gateway doesn't give you generally available tracing.
How to read a GA with Beta inside it #
GA on the control plane means the routing, the provider integration, the grants and the usage monitoring are things Databricks will support in production. Build on those. The Beta capabilities are worth enabling in a non-production workspace now, because they're where the interesting enforcement lives, and because enabling them later is an account-level change rather than a workspace one.
What this should change in a procurement conversation is the question. "Is Unity Gateway GA" has a yes answer that tells you little. Ask instead which of the specific controls in your requirements list is GA today, and get the stage of each one by name, from budgets through service policies, agent services, unified tracing and the managed MCP connectors. Databricks publishes the stage per capability in the Unity Gateway release notes, and that page is a better source for a risk register than any announcement.
The gateway also now carries a long and growing model list through Foundation Model APIs, which is the part teams notice first and the part that matters least to the governance argument. Switching models without changing application code is useful. Knowing which policy actually blocked a call, and being able to prove it to an auditor, is the thing you're buying, and that proof still depends on Beta pieces. Unity Catalog covers the grant model the gateway inherits, and Foundation Models and the AI Gateway covers the serving side.
Questions people ask #
Is Databricks Unity Gateway generally available? #
Yes. Unity Gateway became generally available on 4 August 2026. Databricks noted at the same time that some capabilities, including service policies and agent services, remain in Beta.
What does Unity Gateway do? #
Databricks lists four functions for it. The gateway controls which AI services teams can use, routes and manages AI traffic across providers, governs MCP servers for access and cost, and monitors usage, cost, access and lineage from one place. It is part of Unity Catalog.
Are Unity Gateway budgets generally available? #
Yes. Budgets, including budgets for Unity Gateway, went generally available on 6 July 2026, with per-user thresholds, overrides, and configurable actions such as email alerting and usage blocking.