Claude on Google Cloud: operators, access and migration limits
Anthropic launched a new way to use Claude through Google Cloud on October 22, 2026. Anthropic operates the API, while Google Cloud handles identity and Marketplace billing. The release notes also announce Opus 5.5 fast mode on this route, with Messages API support documented separately. The change added a delivery option, not a new Claude model.
Two Google Cloud routes with different operators
The official comparison distinguishes two operators. Both routes use Google credentials and appear on a Google Cloud bill, but they are not interchangeable integrations.
| Integration check | Anthropic managed | Google managed |
|---|---|---|
| Operator and limits | Anthropic operates the Claude API and manages rate limits | Google operates Claude on Agent Platform and manages rate limits |
| SDK client | AnthropicGoogleCloud | AnthropicVertex |
| Request destination | claude.googleapis.com; workspace-scoped calls include a workspace in the URL | Agent Platform prediction endpoints; model in the URL |
| Streaming | Request-body option | Separate streaming endpoint |
Direct calls to Anthropic with an API key are neither Google Cloud offering. Identify the client and endpoint before planning a migration.
Migration creates a separate organization
The setup and migration guide requires a Marketplace subscription, enabled services and a new Anthropic organization. Existing files, skills, batches, workspaces and agent resources do not transfer automatically. An existing zero data retention arrangement does not carry over either. Teams can run old and new organizations in parallel while moving workloads.
API keys, MCP tunnels, memory stores in environments hosted by the customer and prebuilt sandbox workers hosted by the customer are unsupported here. Custom workers can use the documented work endpoints. Anthropic is the data processor, and some supporting services process data outside Google Cloud. A Google Cloud bill therefore does not establish that all processing stays there.
Access removal is not an agent stop button
Revoking a creator’s Google Cloud access does not stop existing Managed Agents sessions or scheduled deployments. Interrupt running sessions and separately pause, update or archive deployments. Review those resources when removing a user or rotating a service account.
The IAM reference distinguishes three roles. Viewer can inspect workspace resources but cannot send messages or count tokens. User adds workspace operations, including model calls and worker processing. Admin adds permissions over the Anthropic organization. Roles are granted on the Google Cloud project.
Viewer access still deserves scrutiny. It allows downloading files and batch results and reading complete session event histories. A role that cannot launch a job can nevertheless read its contents. Review the information a team needs, not just whether it can spend money by making requests.
Fast mode changes the cost calculation
Anthropic’s fast mode documentation lists Opus 5.5 fast mode at $8 USD per million input tokens and $40 USD per million output tokens as of November 2, 2026. It describes the same model with a faster inference configuration, not greater intelligence. The advertised improvement is up to 2.5 times the output token rate, not an equivalent reduction in time before the first token.
Fast mode has its own rate limit. Caching and data residency modifiers still apply. For an interactive application, compare the value of faster streamed output against those charges rather than assuming every request becomes proportionally faster. These are documented vendor specifications, not ByteForward benchmark results.
Review stored data before moving an agent
The data retention guide makes another distinction important for agent workloads. Managed Agents sessions are stateful and their transcripts persist until deleted. The Files API retains files until deletion or configured expiry. Neither feature is eligible for zero data retention. A successful request does not prove that an application remains within a zero data retention arrangement.
The practical migration review is therefore broader than changing an SDK client. Identify the operator and endpoint, check reader permissions, inventory persistent resources and rehearse how ongoing work will be stopped. For related controls, see ByteForward’s reporting on Claude organization web fetch restrictions.
