Databricks AI Gateway: Mosaic AI Gateway is now Unity AI Gateway

The name changed, and that is the first thing to know

Databricks shipped this as Mosaic AI Gateway, part of the Mosaic AI product family it inherited from the MosaicML acquisition. In 2026 it moved onto Unity Catalog as the governance substrate and took the Unity name with it. General availability was announced on 4 August 2026.

Practically: your notes, your vendor deck, and half the blog posts on the internet still say Mosaic AI Gateway. Databricks' own documentation tree retains legacy pages under the old name. They refer to the same gateway. If you are evaluating against a 2025 comparison table, assume the feature list has moved on.

What it actually governs

Databricks positions this as governance for runtime interactions, not just for data assets. The registerable surface is broad.

• Databricks-hosted foundation model serving endpoints
• External models from other providers, proxied through the gateway
• AI agents, including coding agents
• MCP servers and the tools they expose
• Skills

Each is registered as a Unity Catalog securable, which is what makes the permissions, lineage, and audit story consistent with the rest of the Databricks estate. If your organisation already runs Unity Catalog for table-level access control, this is the same grant model applied to models and tools.

Native multi-model access covers Anthropic, OpenAI, Gemini, Kimi, GLM, and others. Databricks cites over a quadrillion tokens through the gateway in the past year and a single customer at more than 100 billion tokens a month, which is the scale claim rather than a limit.

How it bills, and why nobody quotes a price

Every gateway article wants a dollar figure and this one does not have one, because Databricks does not bill AI Gateway as a line item. Consumption normalises into Databricks Units, and the gateway's costs show up inside your existing DBU bill alongside model serving.

Cost breaks down by model service (usage_metadata.ai_gateway.endpoint_name), by target model (destination_model), by requesting principal (identity_metadata.run_by, user or service principal), and by custom tags such as team or cost centre.

Two consumption modes exist underneath: provisioned throughput, where you commit to reserved capacity and pay for it around the clock, and pay-per-token, which is the usual choice for development and intermittent workloads. Provisioned throughput is where a Databricks AI bill gets large quietly, because reserved capacity bills whether or not you send traffic.

Putting controls on an endpoint

The gateway is configured per endpoint rather than globally, which is the right granularity but means a policy is only as good as your coverage of the endpoints people actually call. The order below is the one that fails least often.

SELECT
  usage_metadata.ai_gateway.endpoint_name    AS service,
  usage_metadata.ai_gateway.destination_model AS model,
  identity_metadata.run_by                    AS principal,
  SUM(estimated_usd)                          AS spend_usd
FROM system.ai_gateway.external_model_spend
WHERE usage_date >= current_date() - INTERVAL 30 DAYS
GROUP BY ALL
ORDER BY spend_usd DESC;

Column names vary by cloud and release, so treat that query as the shape rather than as copy-paste. The point is that attribution is a table you can join against, not a screenshot: the gateway's value in a large organisation is that chargeback becomes SQL.

Who it is for

It fits when

• You already run Unity Catalog and want one permission model for tables, models, agents, and tools.
• Governance is the requirement and cost control is a benefit, rather than the other way round.
• You need request and response payloads landing in Delta tables you own, for audit or evaluation.
• Multiple teams share model endpoints and you need per-principal attribution and hard budget caps.

It does not fit when

• You are not a Databricks shop. There is no standalone Unity AI Gateway. Adopting it means adopting the platform.
• You want a fast, cheap proxy in front of two providers. LiteLLM, Bifrost, or Cloudflare do that in an afternoon and cost nothing.
• You need a published price to put in a business case. DBU-based billing is not quotable without a workload model.
• Your problem is developer laptops. A platform gateway cannot see a coding agent running on a subscription on someone's Mac.

Questions people ask

Keep reading

Continue through LLM gateways and routing with Kong AI Gateway: the AI plugins on top of the gateway you run.