Until now, calling Claude from a Google Cloud project came with fine print: no Files API, no code execution, no Agent Skills, no Message Batches API and no Claude Managed Agents. Since October 22, 2026 there is a second route, already generally available, where Anthropic runs the API inside Google Cloud and those features ship the same day they reach the native platform, billed through Google Cloud and your spend commitments.
Two offerings, one name
Google Cloud keeps the existing option, now called Google managed, and adds Anthropic managed. The models are the same (Claude Fable 5.1, Opus 5.5, Sonnet 5.5 and Haiku 5.5) and both land on your Google Cloud bill. What changes is who operates the service, and with it which features and guarantees you get.
| Anthropic managed (new, GA) | Google managed (existing) | |
|---|---|---|
| Data processor | Anthropic, under its Commercial Terms | Google Cloud, under your current agreement |
| API | The Claude API: claude.googleapis.com/.../invoke/v1/messages | Prediction endpoints: :rawPredict and :streamRawPredict |
| New features | Same day as the Claude API | Once Google integrates them |
| Files API, code execution, Skills, Batches, Managed Agents | Yes | No |
| Location | global; US-only inference with inference_geo: "us" | Global, us/eu multi-region or regional |
| Google integrations | IAM, Cloud Audit Logs, CMEK with Cloud KMS | VPC Service Controls, CMEK, BigQuery, Model Armor |
| Rate limits | Managed by Anthropic | Managed by Google Cloud |
| Billing | Marketplace, draws down eligible commitments | Agent Platform usage, draws down eligible commitments |
The table is the decision. If your team builds agents and wants Managed Agents, Skills or the Files API without leaving Google Cloud billing, the new option is the only way to get them. If you need Google as the data processor, EU data residency or a VPC Service Controls perimeter, you stay on Google managed.
Turning it on
Setup is per project and creates a brand-new Anthropic organization, separate from any your company already has (a Claude Enterprise one included).
- In Google Cloud Marketplace, subscribe to Claude Models on Google Cloud (Anthropic managed) and click Manage on Console.
- Enable both services on every project that will call Claude:
gcloud services enable claude.googleapis.com \
anthropic-managed-claude.cloudpartnerservices.goog \
--project=my-project-id- On the Claude page, click Activate and finish the organization setup. Open Claude Console then takes you to
platform.claude.com, where you copy the workspace ID (it starts withwrkspc_). - Grant
roles/claude.userto the identity that makes the calls:
gcloud projects add-iam-policy-binding my-project-id \
--member="serviceAccount:[email protected]" \
--role="roles/claude.user"In code, the client is AnthropicGoogleCloud (in beta across the SDKs). It picks up Google's Application Default Credentials and reads the project and workspace from environment variables. The rest of the call is exactly what you'd send to the Claude API:
# pip install -U "anthropic[google_cloud]"
# export ANTHROPIC_GOOGLE_CLOUD_PROJECT="my-project-id"
# export ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID="wrkspc_..."
from anthropic import AnthropicGoogleCloud
client = AnthropicGoogleCloud()
message = client.messages.create(
model="claude-opus-5-5",
max_tokens=1024,
messages=[{"role": "user", "content": "Hello, Claude"}],
)Coming from Google managed, the move is mostly a client swap (AnthropicVertex becomes AnthropicGoogleCloud), new IAM roles and a new URL. The model goes back into the request body and streaming is just "stream": true.
What it costs, and the fine print
Per-token pricing is standard Claude pricing, metered in Claude Consumption Units worth $0.01 each, and usage counts toward your Google Cloud commitment like any other Marketplace purchase. If you have a negotiated discount, talk to your account rep before creating the organization: it can't be applied to past usage. Fast mode on Opus 5.5, up to 2.5x faster output, costs $8 per million input tokens and $40 per million output tokens.
A few details matter in production:
- Access is checked at creation, not at runtime. A Managed Agents session or a scheduled deployment keeps running after you revoke its creator's IAM access. When someone leaves or a service account is rotated, review and stop what that identity launched.
- No API keys and no MCP tunnels. Everything goes through Google credentials.
- Roles are granted per project, so they apply to every workspace in it. Teams that need separate access need separate projects, each with its own Anthropic organization.
- Some processing leaves Google Cloud: Anthropic's trust and safety systems handle data outside it. Retention follows the Claude API policy, and zero data retention has to be requested again even if another organization of yours already has it.
My one caution: don't treat this as "Claude on Vertex with extra features". The data processor, data residency and who raises your rate limits all change, and someone from compliance should sign off before production traffic moves.
Sources
- Claude Platform is now available on Google Cloud, Anthropic, October 22, 2026.
- Expanding Claude availability and deployment choice on Google Cloud, Google Cloud Blog.
- Claude Models on Google Cloud, Claude Platform Docs.
- Claude Models on Google Cloud (Anthropic managed), Claude Platform Docs.
- Claude Models on Google Cloud (Google managed), Claude Platform Docs.