Monitor Unity AI Gateway cost
This feature is in Beta.
Observe and analyze cost for all Unity AI Gateway traffic by model service, target model, and requesting principal.
Cost observability is based on Databricks billing records. For request-level usage analytics such as token counts, latency, requester details, and request tags, see Model usage for Unity AI Gateway services.
Requirements
- Unity AI Gateway enabled for your account.
- A Databricks workspace in a Unity AI Gateway supported region.
- The billable usage system table enabled for your account. See Enable system tables.
Attribution
Unity AI Gateway provides cost attribution through the billable usage system table (system.billing.usage).
Unity AI Gateway enriches MODEL_SERVING billing records in system.billing.usage with service-specific metadata, so you can attribute Databricks cost to the associated services, target models, principals, and service tags. For the complete schema and field definitions, see the Billing usage system table reference.
The billable usage system table includes cost attribution for Databricks-hosted models. For external model cost analysis in the dashboard, see External model cost.
For requests served through a Unity AI Gateway model service, Databricks populates the following fields on MODEL_SERVING records in system.billing.usage:
Field | Description |
|---|---|
| The name of the Unity AI Gateway model service that received the request. This is the Unity Catalog fully qualified name, in the form |
| The ID of the Unity AI Gateway model service. |
| The destination model that handled the request, for example |
| The ID of the target that handled the request. |
| The user or Databricks service principal that issued the request. |
| Service tags configured on the Unity AI Gateway model service, such as |
Unity AI Gateway populates these fields for both real-time and batch inference requests routed through it.
Observability
The built-in usage dashboard includes a Cost Analysis page for monitoring cost and analyzing cost breakdowns over time. You can analyze cost across multiple dimensions, including:
- Model service
- Target model
- Requesting user or Databricks service principal
To open the dashboard, click View Dashboard from the AI Gateway page. For details on importing and updating the dashboard, see Built-in usage dashboard.
Cost observability is available in dashboard version 0.4 and above. Account admins must update the dashboard to receive the latest template changes. See Built-in usage dashboard.
External model cost
The usage dashboard can be configured to include cost estimates for external models by specifying a model pricing table in the Pricing Table Override setting. The pricing table is user-managed and must be provided as input to the dashboard.
The pricing table must include the following fields:
Field | Type | Description |
|---|---|---|
| STRING | The model name used for cost attribution in the dashboard. |
| DOUBLE | The price for input tokens. |
| DOUBLE | The price for output tokens. |
| DOUBLE | The price for cache-read input tokens, when supported. |
| DOUBLE | The price for cache-write input tokens, when supported. |
Cost estimates for external models are for informational purposes only. These figures are calculated based on list or override prices and might not reflect your final provider invoice. Databricks is not liable for discrepancies in third-party billing.
Analyzing cost
Genie Code (Agent mode) can do this for you. Try this example prompt:
Query system.billing.usage to show AI Gateway DBU cost for the past 30 days, broken down by usage_metadata.ai_gateway.endpoint_name, destination model, and requesting user. Filter to MODEL_SERVING records. Show top 10 in each.
The following queries analyze cost for Databricks-hosted models in system.billing.usage. Cost can be broken down by model service, target model, and principal.
By model service
SELECT
usage_metadata.ai_gateway.endpoint_name AS endpoint_name,
SUM(usage_quantity) AS dbus
FROM system.billing.usage
WHERE billing_origin_product = 'MODEL_SERVING'
AND usage_metadata.ai_gateway.endpoint_name IS NOT NULL
AND usage_unit = 'DBU'
AND usage_date >= current_date() - INTERVAL 30 DAYS
GROUP BY endpoint_name
ORDER BY dbus DESC;
By destination model
SELECT
usage_metadata.ai_gateway.destination_model AS destination_model,
SUM(usage_quantity) AS dbus
FROM system.billing.usage
WHERE billing_origin_product = 'MODEL_SERVING'
AND usage_metadata.ai_gateway.endpoint_name IS NOT NULL
AND usage_unit = 'DBU'
AND usage_date >= current_date() - INTERVAL 30 DAYS
GROUP BY destination_model
ORDER BY dbus DESC;
By user or Databricks service principal
SELECT
identity_metadata.run_by AS run_by,
SUM(usage_quantity) AS dbus
FROM system.billing.usage
WHERE billing_origin_product = 'MODEL_SERVING'
AND usage_metadata.ai_gateway.endpoint_name IS NOT NULL
AND identity_metadata.run_by IS NOT NULL
AND usage_unit = 'DBU'
AND usage_date >= current_date() - INTERVAL 30 DAYS
GROUP BY run_by
ORDER BY dbus DESC;
Limitations
- Spend attribution applies to
MODEL_SERVINGrecords insystem.billing.usage. Requests routed to external models that are billed directly by the external provider do not appear insystem.billing.usage. - For model services with multiple destinations, such as traffic splitting or fallbacks,
ai_gateway.destination_modelandai_gateway.destination_ididentify the destination that ultimately served the request.