Google Cloud launches Anthropic-managed Claude deployments with Managed Agents, skills and code execution
Google Cloud announced the general availability of Anthropic-managed Claude deployments on October 22, 2026, adding a second operating model for accessing Anthropic frontier models through the Google Cloud ecosystem.
The new option provides access to Anthropic's developer platform, including Claude Managed Agents in beta, reusable skills and code execution, while retaining Google Cloud identity integration, audit logging and consolidated billing.
The established Google-managed Claude deployment remains available for customers prioritizing Google-operated inference, enterprise security boundaries and regional data residency. Both options offer day-zero model access and equivalent standard model pricing, but they do not have identical processing responsibilities, tools or compliance coverage.
··········
WHAT GOOGLE CLOUD ANNOUNCED ON OCTOBER 22, 2026.
Anthropic-managed Claude became generally available as an additional deployment option, not a replacement for the existing Google-managed offering. The general-availability designation applies to the deployment service, not to every individual tool: Managed Agents remains beta.
Anthropic-managed deployments emphasize first-party platform capabilities and immediate access to new Anthropic APIs. Google-managed deployments emphasize Google-operated inference and integrations with enterprise governance services.
........
Characteristic | Anthropic-managed | Google-managed |
Availability | Generally available | Existing production offering |
Inference operator | Anthropic | Google Cloud |
New developer tools | First-party release cadence | Integration-dependent |
Model releases | Day zero | Day zero |
Billing | Cloud Marketplace | Agent Platform |
........
The announcement concerns deployment infrastructure rather than the release of a newly trained Claude foundation model.
··········
TWO OPERATING MODELS WITH DIFFERENT DATA PROCESSORS.
In the Anthropic-managed configuration, Anthropic operates the Claude API and underlying model service. Google Cloud supplies integration functions such as IAM authentication, API routing, Cloud Audit Logs and Marketplace billing.
Anthropic's Commercial Terms, Data Processing Addendum and Usage Policy govern its managed model platform; Google's terms govern the integration layer. The companies act as separate processors for their respective service components.
Under the Google-managed configuration, Google operates inference and serves as data processor under applicable Google Cloud customer agreements.
........
Responsibility | Anthropic-managed | Google-managed |
Model execution | Anthropic | Google Cloud |
Model data processor | Anthropic | Google Cloud |
Model-service terms | Anthropic Commercial Terms | Google Cloud agreements |
Identity | Google IAM | Google IAM |
Quotas | Anthropic | Google Cloud |
........
Using Google Cloud billing does not mean all inference and associated processing remain within Google's infrastructure.
··········
HOW THE ANTHROPIC-MANAGED ARCHITECTURE WORKS.
The Anthropic-managed API is exposed through claude.googleapis.com and uses the Claude Messages API /v1/messages interface. Applications authenticate with Google Cloud IAM, including Application Default Credentials, service accounts or Google access tokens.
Google documents the AnthropicGoogleCloud Python client. The Google-managed configuration instead uses regional Agent Platform endpoints under REGION-aiplatform.googleapis.com.
Although the Anthropic-managed endpoint uses a Google-associated domain, inference and storage are operated by Anthropic and may occur outside Google Cloud. Endpoint naming does not establish data-processing location.
Google has not published a controlled latency or throughput comparison between the two paths, so neither can be characterized as universally faster.
··········
CLAUDE MANAGED AGENTS: HOSTED EXECUTION FOR LONG-RUNNING WORK.
Claude Managed Agents, available in beta, provides a managed agent harness for workflows that require more than a single model response. Such workflows may plan tasks, call tools, work with files, execute code and continue across multiple steps.
The managed runtime reduces the infrastructure developers must build for orchestration, execution environments and task lifecycle management. It is particularly relevant to asynchronous and long-running tasks.
Managed execution does not remove the need for permissions, approval boundaries, tool restrictions, retry policies and recovery from partially completed operations.
The beta status means supported behaviors and limits may change; the announcement does not include a standardized comparison against self-hosted agent frameworks.
··········
CODE EXECUTION, SKILLS AND OTHER DEVELOPER TOOLS.
Anthropic-managed Claude exposes code execution in a managed sandbox, enabling programmatic calculations, file transformation, data analysis and iterative tool-assisted workflows.
Reusable skills package task-specific instructions and resources so agents can follow consistent procedures without retraining the foundation model. External systems still require authorized integrations and appropriately scoped permissions.
Other capabilities include file reuse, computer use, web search and fetch, and remote MCP integrations, subject to their individual availability and release status.
........
Capability | Function | Consideration |
Managed Agents beta | Hosted orchestration | Beta lifecycle |
Code execution | Sandboxed programs | Execution permissions |
Skills | Reusable workflows | Version governance |
Files API | File reuse | Retention policies |
Remote MCP beta | External tools | Authorization |
Web search and fetch | External retrieval | Content trust |
........
A code-execution sandbox is not unrestricted access to customer infrastructure. Security depends on configured tool boundaries and permitted data flows.
··········
MODEL RELEASE PARITY DOES NOT MEAN FEATURE PARITY.
Google states that both configurations provide day-zero access to newly released Claude models. However, first-party Anthropic tools can appear immediately in Anthropic-managed deployments while Google-managed integrations follow a separate validation and release schedule.
Anthropic-managed deployments can pass supported beta feature headers through to Anthropic. Google-managed deployments offer a different set of integrated tools, quotas and service tiers.
........
Feature | Anthropic-managed | Google-managed |
New models | Day zero | Day zero |
First-party tools | Immediate where supported | Integration-dependent |
Beta headers | Supported passthrough | Varies |
Managed Agents | Beta access | Not equivalent parity |
Quota management | Anthropic | Google Cloud |
........
Applications migrating between endpoints should test tool support, authentication, quota behavior and operational constraints instead of assuming drop-in equivalence.
··········
SECURITY: IAM AND AUDIT LOGS ARE NOT IDENTICAL INFERENCE BOUNDARIES.
Both options use Google Cloud IAM. The Anthropic-managed service also provides Cloud Audit Logging for Google's integration layer, but these controls do not automatically extend Google-managed inference protections to Anthropic's service.
Customers requiring Google-operated inference, VPC Service Controls, Customer-Managed Encryption Keys or particular Google Cloud compliance-program coverage should evaluate the Google-managed option.
Compliance programs including FedRAMP High, IL4, IL5, SOC, ISO and HIPAA eligibility are subject to specific service scopes and contracts; they cannot be assumed to cover Anthropic-managed processing solely through Google Cloud authentication.
........
Security issue | Anthropic-managed | Google-managed |
Google IAM | Supported | Supported |
Audit logs | Integration layer | Google-managed service |
Google-operated inference | No | Yes |
VPC Service Controls | Not inherited | Supported |
CMEK | Not inherited | Supported |
Compliance programs | Separate assessment | Eligible service scope |
........
Security assessment must follow the actual model-serving boundary, processing terms, subprocessors and data flows.
··········
DATA RESIDENCY, INFERENCE GEOGRAPHY AND RETENTION.
Google-managed Claude supports regional residency configurations in the United States, European Union and Asia-Pacific, subject to service availability and settings.
Anthropic-managed inference operates on Anthropic-managed infrastructure, and processing may take place outside Google Cloud. Its inference_geo parameter can specify an inference geography for a request, but this is not proof that every log, artifact, subprocess or storage operation remains in the same location.
Anthropic-managed data retention follows applicable Anthropic first-party API terms. Zero Data Retention may be available on request, rather than being automatically enabled for every customer.
........
Governance | Anthropic-managed | Google-managed |
Operator | Anthropic | Google Cloud |
Geography | inference_geo option | Regional controls |
Storage | Anthropic infrastructure | Google infrastructure |
Retention | Anthropic API terms | Google service terms |
ZDR | May be requested | Service-dependent |
........
European and regulated customers should review the exact scope of geographic controls, contractual safeguards and subprocessors before processing sensitive information.
··········
BILLING, PRICING AND GOOGLE CLOUD COMMITMENTS.
Google announced equivalent standard model pricing across the two configurations. Anthropic-managed Claude bills through Google Cloud Marketplace; Google-managed Claude uses Google Cloud Agent Platform billing.
Eligible Google Cloud spending commitments may apply under the relevant Marketplace drawdown policy, but not every commitment necessarily qualifies.
........
Commercial item | Anthropic-managed | Google-managed |
Standard model pricing | Equivalent | Equivalent |
Billing | Cloud Marketplace | Agent Platform |
Commitments | Eligible policy applies | Applicable terms |
Agent tools | Additional costs possible | Additional costs possible |
Task-level cost parity | Not established | Not established |
........
Total application costs can differ even when standard model token prices match. Agent runtime, tool execution, code execution, caching, storage and service tiers may create additional charges.
The announcement provides no universal per-completed-agent-task price, total-cost parity guarantee or comparative infrastructure spending benchmark.
··········
QUOTAS, SERVICE TIERS AND OPERATIONAL SUPPORT.
Anthropic administers rate limits and quotas for its managed deployment, while Google Cloud manages quotas for Google-managed Claude through Agent Platform.
Anthropic-managed tiers include supported Priority, Standard and Batch arrangements. Google-managed options include Google-specific capacity mechanisms such as Provisioned Throughput and batch processing.
Google provides frontline support for integration issues, with Anthropic handling specialist issues involving its own managed model service.
Existing quota allocations should not be assumed to transfer between deployments, and the announcement does not establish equivalent service-level agreements or incident-resolution performance.
··········
BIGQUERY, MODEL ARMOR AND GOOGLE AGENT PLATFORM INTEGRATIONS.
Google-managed Claude retains native integrations with Google Cloud services including BigQuery, Model Armor and Agent Engine, supporting organizations that already rely on Google's data, security and agent infrastructure.
Anthropic-managed Claude emphasizes direct access to Anthropic's developer-platform tools and their release cadence. An organization can use different deployments for different workloads instead of selecting one for every application.
A regulated application may favor Google-managed inference, while an experimental agent application may benefit from the faster availability of Anthropic tools.
··········
CHOOSING THE APPROPRIATE ENTERPRISE DEPLOYMENT.
The decision should begin with the workload's processing boundary, security requirements, tool needs, geographic restrictions, quotas and operational ownership rather than the model name alone.
........
Requirement | Starting choice | Reason |
Newest Anthropic tools | Anthropic-managed | Feature availability |
Managed Agents beta | Anthropic-managed | Hosted harness |
Google-operated inference | Google-managed | Serving control |
VPC/CMEK governance | Google-managed | Google security boundary |
BigQuery and Model Armor | Google-managed | Native integration |
Google billing | Either | Both supported |
........
The two paths can expose the same Claude model while delivering different tool capabilities and enterprise controls. Differences in application outcomes may reflect configuration or orchestration rather than the underlying model.
··········
LIMITATIONS AND WHAT THE ANNOUNCEMENT DOES NOT ESTABLISH.
General availability of Anthropic-managed deployment does not make all Anthropic tools generally available; Managed Agents is still beta. Neither model-price parity nor common Google billing establishes identical total cost.
The two deployments have different processors, service terms, data residency options, security boundaries and quota administration. Google does not publish controlled comparative performance benchmarks or guarantee immediate parity for every developer tool.
The documented inference_geo setting should not be interpreted as a universal residency guarantee for every data flow. Organizations should evaluate current technical documentation and contracts for their intended workload.
··········
WHAT THE NEW DEPLOYMENT MODEL MEANS FOR ENTERPRISE AI.
Google Cloud's October 22 release separates two competing priorities: rapid access to Anthropic's evolving developer platform and tighter control over Google-operated enterprise inference.
Anthropic-managed Claude combines first-party agent tools with Google identity and billing, while Google-managed Claude retains the governance and integration properties of Google's cloud environment.
The same model family can therefore be deployed under materially different processing, security and operational arrangements. Selecting the appropriate configuration requires evaluating the entire service boundary, not only API access or headline token pricing.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
