Google Cloud expands Claude support on Agent Platform

Google Cloud has expanded support for Anthropic's Claude models on its Agent Platform, targeting organisations running AI workloads in production.

The service makes Claude available through Model Garden as a managed offering on Google Cloud, with access over standard REST and JSON interfaces. Customers can use the same Identity and Access Management policies, VPC controls, and logging and monitoring tools they already use across other cloud services.

The move reflects growing competition among large cloud providers to host third-party foundation models while reducing the operational burden on customers. Running advanced AI systems at scale has become a central concern for companies seeking consistent latency, regional data controls, and reliable handling of large requests.

Managed service

The platform handles compute provisioning, auto-scaling, load balancing, and failover for Claude workloads. As a result, customers do not need to build and maintain their own inference clusters to deploy the models in production.

Claude can be accessed through the AnthropicVertex client, which supports prompt caching, tool use, structured outputs, streaming, and adaptive thinking. For batch workloads, customers can use Vertex AI Batch Prediction.

Authentication runs through Application Default Credentials, while requests inherit project-level IAM and VPC settings. Google is positioning this as a way to fold model access into existing cloud governance and operations processes, rather than requiring separate credentials and controls.

Global routing

A major part of the announcement centres on endpoint options for customers with users in different regions. Agent Platform offers global, regional, and multi-region endpoints for Claude, each designed for different production requirements.

Global endpoints route requests to regions where AI compute capacity is available. This allows traffic to shift automatically if one region is constrained, giving customers failover and geographic load balancing without requiring routing logic in their own applications.

Regional endpoints are designed for workloads that require prompts, completions, and intermediate data to remain within a specific geography. Multi-region endpoints are intended to provide US or EU data residency while avoiding dependence on a single region.

These options address a common trade-off in AI deployment. A single endpoint can create latency problems for international users and leave applications exposed to outages or shortages in one location, while duplicating infrastructure across continents is expensive and difficult for many companies.

Compliance focus

Google also highlighted regulated sectors including financial services, healthcare, and government. Claude on Agent Platform inherits Google Cloud's security controls and supports compliance frameworks including FedRAMP High and HIPAA.

VPC Service Controls can be used to define perimeters around Agent Platform resources, which can help prevent data exfiltration. IAM-native access control means customers can govern Claude endpoints with the same roles and policies used for other Google Cloud resources.

Cloud Logging and Cloud Monitoring provide visibility into token usage, error rates, latency, and quota consumption. Combined with regional and multi-region deployment options, these controls give regulated customers a route to using frontier AI systems without redesigning their compliance approach around inference.

Cost and performance

Google said cost and performance remain the main architectural concerns for production AI deployments, and outlined a mix of model-level and infrastructure-level features intended to address both.

Among the Claude features supported on Agent Platform is prompt caching, which can reduce latency by up to 80% and cost by up to 90% by reusing shared prompt prefixes. Streaming responses over server-sent events are also supported for applications such as chat interfaces and coding tools, where response time affects the user experience.

The platform also supports extended and adaptive thinking, allowing the model to adjust how much reasoning it applies to more complex tasks. Google said newer Claude models can support context windows of up to 1 million tokens, enabling longer document analysis, work across larger codebases, and deeper multi-turn conversations.

On the infrastructure side, Agent Platform offers batch prediction for large-scale offline work at lower cost and provisioned throughput for customers that want reserved inference capacity for critical workloads. Google added that memory management and scheduling for long-context requests are handled at the infrastructure layer.

Agent use

Google also linked the Claude deployment model to its broader Agent Platform strategy. The same infrastructure used for inference also supports the agent layer, with development options across Python, Go, Java, and TypeScript, and deployment choices including Agent Runtime, Cloud Run, and Google Kubernetes Engine.

Google added that agents built around Claude can interoperate through the Agent2Agent protocol, which it said is in use at more than 150 organisations. "The result: a planning agent built on Claude can orchestrate sub-tasks across the broader agent ecosystem, under unified IAM, fully auditable, on the same infrastructure that serves the underlying inference."