Claude on Vertex AI is reachable through three kinds of endpoints, and the choice is encoded directly in the URL your requests hit (and in the region argument of the AnthropicVertex client). Getting this right up front matters because endpoint choice is entangled with pricing, quota, provisioned throughput eligibility, and residency — not just latency.
The three endpoint types
| Type | Routing behavior | Client setting | Pricing |
|---|---|---|---|
| Global | Dynamic routing to any region the model supports — maximum availability | region="global", host aiplatform.googleapis.com | List price |
| Multi-region | Dynamic routing within one geography (currently us and eu) — residency with high availability | region="us" or "eu", hosts aiplatform.us.rep.googleapis.com / aiplatform.eu.rep.googleapis.com | +10% over global |
| Regional | Guaranteed routing through one specific region, e.g. us-east1 or europe-west1 | region="us-east1" etc., host LOCATION-aiplatform.googleapis.com | +10% over global |
The 10% regional and multi-region premium applies to Claude Sonnet 4.5 and later models. On the global endpoint, Google states it "can process and serve your requests from any region that is supported by the model" — which improves availability and reduces errors, possibly at the cost of somewhat higher latency for some calls.
When the global endpoint is right
Default to global when your priority is throughput and uptime rather than where processing happens. Availability is its headline feature: capacity pressure in one region is absorbed by routing elsewhere, so you see fewer errors. It is also the cheapest option, and it supports the broad current Claude lineup (Sonnet 5, Fable 5, the Opus 4.x line, Haiku 4.5, and older Sonnets), with prompt caching working on the global endpoint too. Quota is its own independent bucket there — global-endpoint usage does not consume your multi-region quota, and vice versa, which you can exploit deliberately (see quota types).
Know the global endpoint's exclusions before committing: provisioned throughput is not supported (pay-as-you-go only), and Google's batch prediction for Claude does not support the global endpoint either.
When to pin a region — or a geography
Data residency is the usual driver. For partner models, Google stores data at rest within the selected region or multi-region — but note its caveat that "the regionalization of data processing may vary." If your obligation is geographic rather than city-specific ("EU only"), the multi-region eu endpoint gives you the residency boundary while keeping dynamic routing across EU regions for availability. Multi-region supports Claude models version 4.7 and later (Opus 4.7/4.8, Fable 5, and newer). If a regulator or contract requires one named region, use a regional endpoint. Confirm the specifics of your compliance obligations with Google — this guide can't do that for you.
Provisioned throughput is the other hard driver: if you buy reserved capacity, it only works on regional endpoints (see the provisioned throughput guide).
Latency consistency is a softer reason: a regional endpoint guarantees where processing happens, which can make tail latency more predictable for latency-sensitive apps near that region. Measure before paying the 10% premium for this alone.
Enforcing the choice organization-wide
If your policy says "no globally routed AI calls," don't rely on code review. Administrators can block the global endpoint with the org policy constraint constraints/gcp.restrictEndpointUsage, forcing workloads onto regional endpoints. Conversely, if you allow global, document it, because the client-side setting is trivially easy for a developer to change.
us/eu) or a named region; provisioned throughput → regional; everything else → global, and bank the 10%.Where to go next
Endpoint choice determines which quota buckets you draw from — read quota types next, then check the cost side in pricing explained. EU-specific residency planning has its own article: EU residency on Vertex.