Teams arriving at Vertex AI from a traditional ops background often assume they need to wire up their own monitoring before Claude goes to production. In fact, Vertex includes a pre-built model observability dashboard that covers fully managed Model-as-a-Service (MaaS) models — Google's own Gemini models and partner models with managed endpoints, which is exactly the category Claude falls into. Before you build custom dashboards, it is worth understanding what this one already gives you and where its documented edges are.
What the dashboard shows
The dashboard focuses on the operational vital signs of your model traffic:
| Signal | What it tells you |
|---|---|
| Model requests per second (QPS) | Traffic volume and shape — spikes, ramps, daily cycles |
| Token throughput | How many tokens you are pushing through the model over time |
| First token latencies | How long users wait before streaming output begins |
| API error rates | The proportion of calls failing, a leading indicator of quota or capacity trouble |
These are the questions an on-call engineer actually asks — "are we serving traffic, how fast, and is any of it failing?" — answered without writing a single Metrics Explorer query. Note that this is operational telemetry, not content: the dashboard shows counts and latencies, not prompts or completions. For request and response payloads you would use request-response logging to BigQuery, which is a separate, opt-in feature.
Where to find it — and why it might not be there
Google's documentation for Claude on Vertex now lives under the Gemini Enterprise Agent Platform branding (the platform formerly presented as Vertex AI generative AI). To reach the dashboard, open the console's Dashboard page, find the Model observability section, and click Show all metrics.
If you cannot find the section at all, that is expected behavior rather than a broken console: the Model observability section only appears after MaaS API calls exist in the project. A freshly enabled project with no Claude traffic yet shows nothing. Make a few real API calls with AnthropicVertex, wait for the metrics to land, and the section materializes. This trips up teams doing infrastructure-as-code dry runs who want to verify their monitoring story before launch — there is nothing to verify until real traffic exists.
The documented exclusions
Two exclusions matter when you interpret the numbers:
Vertex AI Studio traffic is not included. Google documents that only API calls to a model's endpoint are captured — console usage, such as experimenting in Vertex AI Studio, is not. If your data scientists spent the afternoon in the Studio playground, none of that appears on the dashboard. Conversely, if the dashboard shows fewer requests than your team "knows" it made, check whether some of that activity was console experimentation rather than API traffic.
Self-hosted model metrics are excluded. The dashboard covers fully managed MaaS models only. Claude qualifies, so this rarely bites Claude users directly, but it explains why a mixed estate of MaaS partner models and self-deployed open models will not show up in one place.
Going deeper: Metrics Explorer and alerting
The pre-built dashboard is a read-only starting point. The same telemetry is available as raw metrics in Cloud Monitoring's Metrics Explorer under the endpoint resource ("Agent Platform Endpoint > Prediction"), including prediction/online/error_count, prediction/online/prediction_count, prediction/online/prediction_latencies, and prediction/online/response_count. You can filter by endpoint_id — with the quirk that dots in model version strings are written as the letter p — and aggregate by fields such as response_code.
Because these are ordinary Cloud Monitoring metrics, they plug straight into standard alerting policies. Google's own worked example is a capacity alert: fire when more than 1% of requests return HTTP 429, an early signal that you are exhausting quota and should look at 429 handling or a quota increase. For token-level accounting there is one more wrinkle worth knowing — the console Quota page's token figures can be inaccurate for Claude, a story covered in why the Vertex quota page shows inaccurate token counts.
Where to go next
Pair the dashboard with accurate token accounting via Metrics Explorer, and see the BigQuery request-response log schema when you need payloads rather than counts. The platform overview compares observability stances across all four Claude 3P platforms.