Short version. Anthropic now runs its own Claude API inside your Google Cloud project, and Google Cloud Marketplace does the billing. The docs went up on 22 October and Google's launch post followed on 23 October. There are no API keys, usage is metered in Claude Consumption Units at $0.01 each, and the SDK client is still in beta. It isn't the same product as Claude on Vertex AI, so check which one you're actually calling.
You already pay Google Cloud, and your security team already trusts its IAM. So someone asks the obvious question: can Claude just show up on that bill? Since this week, yes, and the catch is that there are now two ways to do it that look similar and behave differently.
The new one is called Claude Models on Google Cloud (Anthropic managed). Anthropic's release notes for 22 October describe it as "the Claude API operated by Anthropic and reached through your Google Cloud project." Google IAM controls access and Google Cloud Marketplace handles billing. Google's own post, dated 23 October, announces general availability of the same option. The older Vertex-style path, which Google manages, stays.
What changes compared with Claude on Vertex AI
Start with the plumbing, because it decides whether your existing code moves over. The Anthropic-managed option uses the normal Claude API shape: endpoints under /v1/, the model named in the request body, and streaming switched on with "stream": true. The Google-managed option keeps prediction-style endpoints, with the model in the URL and a separate streaming method. If you've written a Vertex client, you can't point it at the new service and hope.
Auth is the part that will trip people. The docs say flatly that the service "doesn't use API keys." You send a Google OAuth token with the cloud-platform scope, usually through Application Default Credentials. The SDK client for it, AnthropicGoogleCloud, ignores ANTHROPIC_API_KEY, ANTHROPIC_AUTH_TOKEN and ANTHROPIC_BASE_URL. Install it with the google_cloud extra. Access is granted through three roles, roles/claude.viewer, roles/claude.user and roles/claude.admin, and they apply per project, so they cover every workspace inside it. Only the admin role can open the Claude Console.
Setup also creates a new Anthropic organization for your Google Cloud project. It's separate from any Anthropic organization your company already has. That matters for finance and for access: your existing workspaces, keys and spend limits don't follow you. Rate limits are managed by Anthropic, and the docs give no numbers, only "contact your account representative." Plan for that conversation if you need volume.
The feature list leans toward the Anthropic-managed side. It supports Managed Agents, the Files API, Message Batches, the Models API and fast mode for Claude Opus 5.5. The Google-managed page lists several of those as unsupported, including the Files API, code execution, the MCP connector and Batches. On the other hand, the Anthropic-managed docs list what it lacks: API keys, MCP tunnels, and memory stores on self-hosted environments. If you rely on one of the missing items, that settles it. We covered the fast-mode economics separately in our Opus 5.5 fast mode piece.
Billing: $0.01 units and a 1.1x multiplier
Anthropic meters usage and reports it to Google in Claude Consumption Units. One unit is $0.01 USD. The charges land on your Google Cloud bill and Anthropic doesn't send a separate invoice. If you have a Google Cloud spend commitment, usage counts toward it like other Marketplace purchases, and Google's post says the same: standard pricing that draws down eligible commitments.
That's a real selling point for anyone sitting on a committed spend they need to burn. It's also where I'd slow down. The pages we read don't give a per-model table that turns units into tokens, and the post only says pricing matches across both options. So we can't tell you what a million Sonnet 5.5 output tokens costs on this path, and we won't guess. Compare against the LLM API pricing tracker for list prices, then confirm the unit math on your first invoice.
One number is explicit. US-only inference, set per request with inference_geo: "us", costs 1.1 times the standard rate on input, output, cache writes and cache reads. The location in the URL stays global; the docs say it's separate from where inference runs. Workspace geo is fixed at us. Zero Data Retention is available on request, and Anthropic is the data processor, with some supporting services such as trust and safety systems running outside Google Cloud. If your compliance review assumed everything stays inside Google's perimeter, that last sentence is the one to read twice.
Who should move, and who should wait
If you're a Google Cloud shop with a commitment and you want the full Claude API surface (agents, files, batches) without a second vendor invoice, this is built for you. The model IDs are the same ones the Claude API uses, and the docs use claude-opus-5-5 throughout. Google's post lists Claude Fable 5.1, Sonnet 5.5, Haiku 5.5 and Opus 5.5 as available on both offerings. The Anthropic docs don't enumerate models for this option, though, and an unavailable one returns an error saying it isn't available for your organization's inference configuration.
If you need regional or multi-region endpoints, customer-managed keys under Google's control, or Vertex-style quotas, the Google-managed option is still the better fit. The docs say its multi-region and regional endpoints cost 10 percent more than global ones, which is its own pricing wrinkle. The two sources also differ on detail. Google's post ties some security controls to the Google-managed option, while Anthropic's docs say CMEK with Google Cloud KMS keys is available on its side. I'd verify that with both vendors before a security review leans on it.
My read: this is a good move for Anthropic and a mildly confusing one for buyers. Two products, one name, different auth and different limits. The beta SDK client is the other flag. I'd run a pilot on a non-critical workload first, watch one full billing cycle, and only then migrate anything with an SLA attached. I might be wrong about the pace, since GA is the word Google used, but "GA" and "SDK in beta" in the same launch is worth a pause.
FAQ
Can I use my existing Anthropic API key?
No. The service doesn't use API keys. You authenticate with a Google OAuth token, and the SDK client ignores the usual Anthropic environment variables.
Is this the same as Claude on Vertex AI?
No. Vertex-style access is the Google-managed option, with prediction endpoints and Google-set rate limits. The Anthropic-managed option uses the Claude API shape, its own Anthropic organization and Anthropic-set limits.
How is it billed?
Through Google Cloud Marketplace, in Claude Consumption Units of $0.01 each, on your Google Cloud invoice. Eligible Google Cloud commitments are drawn down. US-only inference costs 1.1 times the standard rate.
Is fast mode available?
Yes, for Claude Opus 5.5, according to Anthropic's 22 October release note.
What isn't published yet?
A unit-to-token table per model, numeric rate limits, and a full model list for this option. We didn't find them in the docs pages we read.
Sources
