When your application calls Claude on Vertex AI, requests go to a Vertex endpoint, not to Anthropic — Claude runs as a fully managed partner model inside Google Cloud. That already answers one residency question (the traffic stays within Google's platform, and Google states customer prompts and responses are not shared with third parties). The remaining question is where within Google Cloud the work happens, and that is decided by which of the three endpoint types you call.
Global, multi-region, and regional
| Endpoint type | Routing behavior | Residency story | Pricing |
|---|---|---|---|
| Global | Any supported region worldwide, for maximum availability | None — processing may occur in any supported region | List price |
Multi-region (eu) | Dynamic routing within a geography | Stays within the EU geography, with high availability | +10% over global |
Regional (e.g. europe-west1) | Guaranteed routing through one specific region | Pinned to that region | +10% over global |
With the global endpoint, Google says plainly that it can process and serve your requests from any region the model supports — great for availability, unusable for an EU-only mandate. The eu multi-region endpoint is usually the sweet spot for European workloads: it keeps dynamic routing (so a single region's capacity constraints don't become your outage) while confining it to one geography. It has its own hostname, aiplatform.eu.rep.googleapis.com, and supports Claude models version 4.7 and later — claude-opus-4-7, claude-opus-4-8, claude-fable-5, and peers. A regional endpoint pins processing to one named region; it is also the only endpoint type that supports provisioned throughput, since global and multi-region are pay-as-you-go only. Which specific EU regions currently offer which Claude model is a moving list — verify your target region on Google's Claude model pages rather than assuming (Anthropic's docs use europe-west1 as an EU regional example in the SDK).
In the Python SDK the choice is one parameter:
from anthropic import AnthropicVertex
# EU multi-region: dynamic routing, confined to the EU geography
client = AnthropicVertex(project_id="my-project", region="eu")
# Or pin to a single EU region
client = AnthropicVertex(project_id="my-project", region="europe-west1")What the guarantee covers — read this part twice
Google's partner-model documentation states that data at rest is stored within the selected region or multi-region, but adds that "the regionalization of data processing may vary." That nuance matters in a compliance review: the storage commitment and the processing commitment are not the same sentence, and the regional endpoint's core promise is about routing your requests through the chosen region. If your obligations hinge on precise processing-location language, get the current wording from Google's documentation and your Google Cloud agreement — this guide can flag the distinction, not resolve it. Claude on Vertex inherits your cloud provider's compliance posture; confirm specifics with your provider.
Also be clear about what an EU endpoint does not restrict: it does not control where your callers are, does not stop another team from instantiating a global client tomorrow, and does not constrain the storage services around the call (your logs, buckets, and BigQuery datasets have their own locations — see log exports and CMEK).
Making EU-only the enforced default, not a convention
- Block the global endpoint by policy. Admins can set the org policy constraint
constraints/gcp.restrictEndpointUsageto block global endpoint usage and force regional endpoints — turning "please use EU" into "only EU works". - Budget the premium. Regional and multi-region endpoints carry a 10% premium over global list prices for Claude Sonnet 4.5 and later (Opus 4.8, for example, is $5.00/$25.00 per million tokens on global and $5.50/$27.50 on multi-region).
- Remember quotas are per endpoint bucket. Global-endpoint quota and each multi-region-endpoint quota are independent; moving traffic to
eumeans requesting capacity there, not reusing your global headroom. - Batch caveat. Google's Vertex batch prediction for Claude does not support the global endpoint — consistent with a residency posture, but plan batch region choices explicitly.
Where to go next
Compare endpoint mechanics in global vs regional endpoints, add network-level control with Private Service Connect, and see data residency across platforms for the cross-cloud view.