How to Use Claude on Vertex AI: Pricing, Claude Code Setup, Quotas, and Google-Managed vs. Anthropic-Managed

There are two ways to use Claude while keeping your Google Cloud contract and billing. One is the Google-managed version, which Google runs as part of Agent Platform (formerly Vertex AI). The other is the Anthropic-managed version, where you use the Claude API run by Anthropic from a Google Cloud project. It became generally available on October 22, 2026 (Anthropic's announcement). Both show up on your Google Cloud bill, but they differ in available features, regions, who manages usage limits, and Claude Code support.

Based on the official Anthropic and Google Cloud documentation (checked October 26, 2026), this article explains how the two versions differ, pricing, regions, how to get started, Claude Code setup, and how to handle quotas and 429 errors. In sections that cover both versions, they appear in the order Google-managed → Anthropic-managed, using the same colors.

1. How the two versions differ

Both are named "Claude Models on Google Cloud," distinguished by "Google managed" or "Anthropic managed" in parentheses. Google now calls Vertex AI "Gemini Enterprise Agent Platform," and its documentation titles have changed to Agent Platform. Claude Code's login screen, however, still shows "Google Vertex AI."

Sources: the comparison table in Anthropic's "Claude Models on Google Cloud," and Google Cloud's "Use Claude models on Google Cloud (Anthropic managed)" and "Anthropic Claude models." Checked October 26, 2026.

They also have things in common. Both are called with Google Cloud credentials (including Application Default Credentials, or ADC), and charges appear on your Google Cloud bill. For Claude 4.6 and later, model IDs are the same as in the Claude API for both versions (for example, claude-opus-5-5). For earlier models, only the Google-managed version puts @ before the date (for example, claude-haiku-4-5@20251001).

With the Anthropic-managed version, a new Anthropic organization is created for each Google Cloud project. It is separate from any organization you already use with the Claude API or from Claude Enterprise, and an existing organization cannot be moved over. API keys, workspaces and Console settings do not carry over, and resources such as files and batches stay in the original organization.

2. How to choose: 5 questions

Go from the top and find the first one that applies. If none apply, decide on how easy the costs are to estimate. The Google-managed version publishes its price list and default quotas, so you can estimate budget and limits before you start. The Anthropic-managed version's breadth of features and early access to new ones become the advantage if you can confirm rates on the Marketplace listing and limits with your Anthropic representative.

• You want to use Claude Code billed through Google Cloud→Google-managed
• You want only Google to handle your data, or you need VPC Service Controls, regional data storage, or certifications such as FedRAMP High→Google-managed
• You need inference to run only inside the EU→Google-managed (eu multi-region)
• You use the Files API, Agent Skills, Managed Agents, web fetch, the MCP connector, code execution, or the Message Batches API→Anthropic-managed
• You want fast mode for Opus 5.5→Anthropic-managed

The second question lists the cases where the Google Cloud documentation recommends choosing the Google-managed version (Use Claude models on Google Cloud (Anthropic managed)).

You can also use both. For example, developers use Claude Code through the Google-managed version, while your own application calls Managed Agents through the Anthropic-managed version. Both appear on the same Google Cloud bill, but the admin consoles and usage-limit management are separate.

3. Feature differences

The biggest differences are in the server-side features that Anthropic operates. The Google-managed version centers on the Messages API, and file storage and agent runtime environments are not available.

Sources: Feature support in Anthropic's "Claude Models on Google Cloud (Anthropic managed)," "Claude Models on Google Cloud (Google managed)," and "Fast mode." Fast mode on the Google-managed version is absent from the supported feature list in Anthropic's docs, which say neither that it is nor that it is not available (Claude Code's feature table lists it as unavailable on Agent Platform). Checked October 26, 2026.

For how Managed Agents work, see "What Are Claude Managed Agents?"; for fast mode's speed and pricing, see "Claude Opus 5.5 release overview."

4. Pricing: is it the same as the Claude API?

Google-managed pricing

On the Google-managed version's global endpoint, input, output and cache rates for Opus 5.5, Sonnet 5.5 and Haiku 5.5 are the same as Anthropic's Claude API. The us and eu multi-regions cost 10% more than global (regional endpoints are also 10% more, but they only serve Sonnet 4.6 and earlier models; see section 5).

Sources: Anthropic "Pricing," Anthropic's Claude models in Google Cloud's "Agent Platform pricing," and Anthropic's "Claude Models on Google Cloud (Google managed)" (multi-region and regional endpoints are 10% more than global, for Sonnet 4.5, Haiku 4.5, Opus 4.5 and later models). Checked October 26, 2026. The Anthropic-managed version is not in the chart because its per-token rates are not in the official docs (see the section below).

US dollars per million tokens. Rates for Anthropic's Claude API and the Google-managed global endpoint (both companies' pricing pages, checked October 26, 2026). The us/eu multi-regions are 1.1x these rates. Cache reads cost 0.05x input for Opus 5.5 and Sonnet 5.5, and 0.1x for Haiku 5.5 (note on Anthropic's pricing page). For the full model list and rate history, see "Claude pricing compared by model."

Provisioned Throughput, which reserves capacity for a fixed fee, is not available on global or the multi-regions and requires a regional endpoint (Google Cloud documentation). Regional endpoints only serve Sonnet 4.6 and earlier models, so Opus 5.5, Sonnet 5.5 and Haiku 5.5 are used pay-as-you-go.

Anthropic-managed pricing

With the Anthropic-managed version, Anthropic meters usage in Claude Consumption Units (CCU, 1 CCU = US$0.01) and reports it to Google Cloud, and it appears on your Google Cloud bill. Anthropic does not send a separate invoice. A per-token price list is not published in the official documentation of either Anthropic or Google Cloud (checked October 26, 2026). The reliable approach is to check rates on the Google Cloud Marketplace listing before subscribing, then once you start, run a small test and confirm the actual amount in the Claude Console's Usage page and Google Cloud's billing reports.

Three things are officially documented:

• Fast mode for Opus 5.5 costs $8 input and $40 output per million tokens (Anthropic's Fast mode page and announcement).
• Adding inference_geo: "us" to a request to keep inference inside the US multiplies prices in every category by 1.1 (Anthropic "Data residency").
• If you have a Google Cloud spend commitment, usage counts toward it like other Marketplace purchases. If you have a discount agreement, contact your Anthropic or Google Cloud representative before you start (discounts are not applied retroactively).

Batch pricing

The two versions handle batches differently. The Google-managed version uses Agent Platform batch prediction (with JSONL input from BigQuery or Cloud Storage), and its rates are in the batch column of Google Cloud's pricing page. The Anthropic-managed version supports the same Message Batches API as the Claude API, but its batch rates are not in the official docs either. The batch discount on the Claude API (direct) is covered in "Claude pricing compared by model."

5. Regions and data storage

For Claude on Google Cloud, the location you call determines where inference runs. The Google-managed version offers three types, while the Anthropic-managed version offers only global.

Sources: Anthropic "Claude Models on Google Cloud (Google managed)" and "(Anthropic managed)," Google Cloud "Agent Platform partner models." Checked October 26, 2026.

Data storage locations also differ. The Google-managed version stores data in the region or multi-region you choose. However, Google Cloud's partner models page also says the processing location may differ from the storage location. Anthropic's documentation, on the other hand, says multi-region inference stays within us or eu, and the relationship between the two statements is not explained. If you have strict requirements on where processing happens, check with Google Cloud before signing a contract. You can also use the organization policy constraints/gcp.restrictEndpointUsage to block the global endpoint and allow only multi-regions.

For the Anthropic-managed version, the workspace storage region is "us" and cannot be changed (Anthropic "Data residency").

As of October 26, 2026, the multi-regions in either version's official documentation cover only the US and the EU. If processing or storage must stay inside a specific country outside those areas, confirm with Google Cloud or your Anthropic representative that a suitable location exists before signing a contract.

6. Getting started

Both versions require a Google Cloud project and a billing account with billing enabled. Below are the steps from the official docs to make each version available, followed by the shared steps for authenticating and calling the model.

Google-managed: getting started

1. Enable the Agent Platform API

   gcloud config set project YOUR-PROJECT-ID
   gcloud services enable aiplatform.googleapis.com

   Do this once per project.
2. Enable the Claude models you want in Model GardenThe person enabling them needs the Consumer Procurement Entitlement Manager role (roles/consumerprocurement.entitlementManager). cloudcommerceconsumerprocurement.googleapis.com must also be allowed by your organization policy. The Claude Code documentation says access approval can take 24 to 48 hours (the Google Cloud documentation does not state how long it takes).
3. Grant a role to callersGrant a role that includes the aiplatform.endpoints.predict permission (roles/aiplatform.user).

Sources: Anthropic "Claude Models on Google Cloud (Google managed)," Google Cloud "Agent Platform partner models." Checked October 26, 2026.

Anthropic-managed: getting started

1. Subscribe in MarketplaceOpen "Claude Models on Google Cloud (Anthropic managed)" in Google Cloud Marketplace and click Subscribe. You need Marketplace purchase permissions on the billing account.
2. Enable two services

   gcloud services enable claude.googleapis.com \
     anthropic-managed-claude.cloudpartnerservices.goog \
     --project=YOUR-PROJECT-ID

   Do this once per project.
3. Activate the organizationClick Activate on the Claude page, then follow the link sent to your contact email to enter your organization name, business type, country and use case, and accept Anthropic's Commercial Terms of Service and Usage Policy. This requires roles/claude.admin.
4. Note your workspace IDIn the Claude Console (opened from the Google Cloud console), go to Settings > Workspaces and note the ID that starts with wrkspc_. You need it to make calls.
5. Grant a role to callersGrant roles/claude.user. Roles are per project, so if you want to separate teams, use separate projects (each project becomes a separate Anthropic organization).

Source: Set up your account in Anthropic's "Claude Models on Google Cloud (Anthropic managed)" (checked October 26, 2026). You cannot sign in to the Anthropic organization with an email address; the Console is opened from the Google Cloud console.

Both versions: authenticate and call

For either version, run gcloud auth application-default login on the calling machine to create ADC, and the SDK will use those credentials (on servers, use service account credentials). Then call the model with the SDK client from section 7.

7. Calling Claude from the SDK

Anthropic's official SDKs support both versions. Only the package to install and the client name differ; messages.create is written the same way as with the Claude API.

Google-managed (Python)

# pip install -U "anthropic[vertex]"
from anthropic import AnthropicVertex

client = AnthropicVertex(project_id="YOUR-PROJECT-ID", region="global")  # "us" or "eu" also work
message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)

In TypeScript, use AnthropicVertex from @anthropic-ai/vertex-sdk. To call over HTTP without the SDK, the URL takes the form https://aiplatform.googleapis.com/v1/projects/YOUR-PROJECT-ID/locations/global/publishers/anthropic/models/claude-opus-5-5:rawPredict; leave the model name out of the request body and put "anthropic_version" (vertex-2023-10-16 or 2026-10-21) in it. For streaming, use :streamRawPredict. What changes in the new API version 2026-10-21 is covered in "temperature is deprecated for this model: causes and fixes."

Anthropic-managed (Python)

# pip install -U "anthropic[google_cloud]"
# Environment variables: ANTHROPIC_GOOGLE_CLOUD_PROJECT and
#           ANTHROPIC_GOOGLE_CLOUD_WORKSPACE_ID (required)
from anthropic import AnthropicGoogleCloud

client = AnthropicGoogleCloud()
message = client.messages.create(
    model="claude-opus-5-5",
    max_tokens=1024,
    messages=[{"role": "user", "content": "Hello, Claude"}],
)

In TypeScript, it is AnthropicGoogleCloud from @anthropic-ai/google-cloud-sdk (both are in beta). Over HTTP, the request and response bodies are the same as the Claude API; the differences are:

• Base URL: https://claude.googleapis.com/v1/projects/{project}/locations/global/workspaces/{workspace}/invoke, which includes the workspace ID (do not send the anthropic-workspace-id header).
• Authentication: send a Google access token with Authorization: Bearer (do not send x-api-key).
• API keys: cannot be created. The ANTHROPIC_API_KEY and ANTHROPIC_BASE_URL environment variables are ignored.

Sources: Anthropic "Claude Models on Google Cloud (Anthropic managed)" and "(Google managed)." Checked October 26, 2026.

8. Using Claude Code on Google Cloud

On Google Cloud, Claude Code supports the Google-managed version. Claude Code's feature table defines Agent Platform as "run by Google, configured with CLAUDE_CODE_USE_VERTEX." The Claude Code documentation does not describe how to use Claude Code with the Anthropic-managed version (checked October 26, 2026).

Two ways to set it up

1. Set it up with the wizardStart claude and, on the login screen, choose "3rd-party platform" → "Google Vertex AI" (if you are already logged in, use /login). It asks for the authentication method, project and region, and lets you check which models are available and pin them. The result is saved in env in your user settings (~/.claude/settings.json), and you can reopen it later with /setup-vertex.
2. Set it up with environment variablesThis suits CI and rolling out to a team. After enabling the models in Model Garden, set the following.

   export CLAUDE_CODE_USE_VERTEX=1
   export CLOUD_ML_REGION=global
   export ANTHROPIC_VERTEX_PROJECT_ID=YOUR-PROJECT-ID
   
   # Pin the models
   # Main model: Sonnet keeps costs down. For Opus, use claude-opus-5-5
   export ANTHROPIC_MODEL=claude-sonnet-5-5
   export ANTHROPIC_DEFAULT_OPUS_MODEL=claude-opus-5-5
   export ANTHROPIC_DEFAULT_SONNET_MODEL=claude-sonnet-5-5
   export ANTHROPIC_DEFAULT_HAIKU_MODEL=claude-haiku-5-5

   The reasons for pinning models are explained in the warning below.

Either way, run /status after setup. If "API provider" shows Google Vertex AI and lines for the GCP project, region and model appear, the setup worked.

Sources: Claude Code "Claude Code on Google Cloud's Agent Platform" and "Environment variables." Checked October 26, 2026.

If a model is not available in the CLOUD_ML_REGION location, you can override the region for that model alone, as in VERTEX_REGION_CLAUDE_HAIKU_4_5=us-east5 (this example calls only Haiku 4.5 in the us-east5 region). CLOUD_ML_REGION takes global, us or eu, or a region name such as us-east5. The Claude Agent SDK also uses the Google-managed version with the same environment variables (see "What Is Claude Agent SDK?").

Claude Code feature support on Agent Platform

Computer use in this table is the Claude Code feature (which requires a subscription), not the API computer use tool in section 3. Source: Claude Code "Feature availability" (/logout and MCP tool search are from "Claude Code on Google Cloud's Agent Platform"). Checked October 26, 2026. claude.ai connectors are not loaded either (see "claude.ai connectors are disabled in Claude Code").

9. Quotas and 429 errors

Google-managed: quotas are per model family

Models released after May 26, 2026 (Opus 5.5, Sonnet 5.5, Haiku 5.5 and others) use quotas shared across the model family (for earlier models, check the per-model quotas on the same Google Cloud page). For example, whichever version of Opus you call on global, it draws from the same anthropic-claude-opus quota, and you don't need to request a quota again when a new version comes out. global and each multi-region have separate quotas.

Source: default quotas in Google Cloud's "Quotas for Anthropic Claude models" (checked October 26, 2026). Input is the total of uncached tokens and cache writes. Google notes that limits vary by account and may be restricted.

Check quotas and request increases on the Quotas & System Limits page in the Google Cloud console. Filtering by base_model (for example, anthropic-claude-opus) makes them easier to find. Token counts on this page may be inaccurate because of how estimation and refunds work; for exact counts, use the token counting API or token_count in Metrics Explorer.

When you exceed a quota, you get a 429 error with this message:

Resource exhausted, please try again later.

This is the message for pay-as-you-go usage (Google Cloud "Error code 429"). With Provisioned Throughput, it is "Too many requests. Exceeded the Provisioned Throughput."

1. Use global if you have no requirement on inference locationIt has larger quotas than the multi-regions and is processed in whichever region has capacity. In Claude Code, set CLOUD_ML_REGION=global. If you have a requirement such as keeping inference inside the EU, skip this and request an increase (step 3) first.
2. Wait and retryRetry with capped exponential backoff, and avoid sudden spikes in volume.
3. Request a quota increaseRequest it from the Quotas & System Limits page mentioned above.

Sources: Google Cloud "Error code 429," troubleshooting in Claude Code's "Claude Code on Google Cloud's Agent Platform."

Anthropic-managed: limits are managed by Anthropic

Usage limits are managed by Anthropic, and default values are not in the official documentation. For increases, talk to your Anthropic representative or sales. Anthropic's documentation recommends confirming your limits before moving production traffic. Errors take the same form as the Claude API: exceeding a limit returns a 429 rate_limit_error, and congestion returns a 529 overloaded_error. If IAM permissions are missing, Google Cloud returns 403.

1. Wait and retryRetry with exponential backoff, and avoid sudden spikes in volume (same as the Google-managed version).
2. Confirm limits before productionBefore migrating, tell your Anthropic representative the volume you need and confirm your limits.
3. Ask for a limit increaseTalk to your Anthropic representative or sales. Limits cannot be raised from Google Cloud's quota page.

Sources: Rate limits in Anthropic's "Claude Models on Google Cloud (Anthropic managed)," and "Errors." Checked October 26, 2026.

Summary

FAQ

Q. Is Claude on Vertex AI more expensive than the Claude API?

On the Google-managed version's global endpoint, input, output and cache rates for Opus 5.5, Sonnet 5.5 and Haiku 5.5 are the same. The us and eu multi-regions are 10% more (Anthropic and Google Cloud pricing pages, checked October 26, 2026). For the Anthropic-managed version, per-token rates are not in the official docs, so check the Google Cloud Marketplace listing.

Q. Can I use a Claude API key?

Both versions are called with Google Cloud authentication (ADC or a service account), not an API key created in the Claude Console. The Anthropic-managed version cannot create API keys, and setting ANTHROPIC_API_KEY is ignored.

Q. Can I use Claude Code with the Anthropic-managed version?

The Claude Code documentation does not describe a setup for the Anthropic-managed version (checked October 26, 2026). Claude Code's Google Cloud setting (CLAUDE_CODE_USE_VERTEX) is for the Google-managed version.

Q. Can I switch my current Anthropic organization to Google Cloud billing?

No. The Anthropic-managed version creates a new organization per project, and existing organizations cannot be converted. API keys, workspaces and Console settings do not carry over, so you recreate them in the new organization.