Latest in AI
Claude on Google Cloud: Anthropic-Managed API, Billing and Migration Limits
The new Google Cloud route brings native Claude features and familiar billing, but changing the operator changes more than the endpoint.
Anthropic made Claude Models on Google Cloud, its Anthropic-managed offering, generally available on October 22. Developers can access native Claude Platform capabilities using Google Cloud credentials and billing. The existing Google-managed offering remains available, so the decision now includes who operates the service and processes the data.
For an enterprise already buying through Google Cloud, that distinction is useful. Consolidating the invoice can simplify procurement, but it does not make two services operationally identical.
What the Anthropic-managed offering adds
The launch announcement says new features and betas arrive when they reach the native Claude Platform. Its examples include Claude Managed Agents in beta, code execution, Skills, web search and prompt caching.
Opus 5.5 fast mode is also available. Anthropic lists input and output prices of $8 and $40 per million tokens for that mode, with an up-to-2.5-times output-speed claim. Those are fast-mode prices, not a single price for every model or workload.
The practical attraction is access to platform features without moving the commercial relationship away from Google Cloud. Teams should still check the specific API they need rather than infer support from the phrase "full platform."
Anthropic managed versus Google managed
Anthropic's comparison documentation describes two different integration paths.
| Detail | Anthropic managed | Google managed |
|---|---|---|
| Service operator | Anthropic | |
| API surface | Claude API endpoints | Agent Platform prediction endpoints |
| SDK client | AnthropicGoogleCloud | AnthropicVertex |
| Billing route | Google Cloud Marketplace | Google Cloud Agent Platform usage |
| Rate limits managed by | Anthropic | Google Cloud |
Both use Google Cloud credentials, but their IAM access is granted separately. The Anthropic-managed path creates a new Anthropic organization for the project. It is not simply the old Google-managed endpoint with a different name.
This makes the migration question more concrete. Compare the features the application uses today, the configuration it depends on and the services its operators monitor. If the only requirement is access to a model already working in production, a second integration path is an option to evaluate rather than an automatic reason to move.
What to check before moving a workload
The migration guide says existing organizations cannot be converted. API keys, workspaces, stored resources and Zero Data Retention arrangements do not carry over.
Authentication uses Google credentials rather than Anthropic API keys. Anthropic is the data processor, and some supporting systems process data outside Google Cloud. These details deserve review separately from the billing decision.
The same guide lists unsupported features: MCP tunnels, memory stores in self-hosted environments and pre-built self-hosted sandbox workers. For the latter, developers can implement a worker against the Environments Work endpoints instead.
Managed Agents also need a careful offboarding process. Revoking a creator's Google Cloud access does not stop an already-running session or scheduled deployment. Those resources need to be stopped separately.
Access management and workload shutdown are separate controls. That is easy to miss if a migration checklist only tests whether a new account can make its first API call. Include the reverse test: what remains running when that identity loses access?
Billing follows the cloud account, not every old agreement
Claude usage appears on the Google Cloud bill and can count toward an existing Google Cloud commitment under the applicable Marketplace rules. Anthropic advises customers with existing discounts to contact their account representative before starting because discounts cannot be applied retroactively.
For a team comparing costs, the useful number is the price of its actual workload under the applicable agreement. Include the selected model, fast mode if used and any separate platform services. A familiar invoice alone does not establish a discount.
The new offering gives Google Cloud customers another way to choose between native Claude capabilities and the Google-managed service. The strongest migration plan will explain why the application needs that choice, then verify the permissions, retained data and running agents that surround the model call.
I'm Paras. I work at the intersection of AI, engineering and distribution.
I help founders explain complex AI infrastructure and enterprise product launches without losing the technical details that buyers care about.