Google Cloud’s AI Powerhouse: Unleashing Anthropic’s Claude Models on Vertex AI

In the rapidly evolving landscape of artificial intelligence, developers are constantly seeking powerful and flexible tools to bring their innovative ideas to life. Google Cloud’s Vertex AI stands at the forefront, providing a comprehensive, fully managed service for deploying and managing advanced AI models, including the powerful family of Anthropic’s Claude models. This integration empowers developers to build and scale AI applications and agents with ease and efficiency, all within Google’s secure and compliant infrastructure.

🎧 If you would prefer to listen instead, you can check out the podcast version of this blog.

The Power of Claude Models on Vertex AI

Anthropic’s Claude models are now accessible on Google Cloud’s Vertex AI as a fully managed service, meaning you can start building without the overhead of infrastructure provisioning or management. Vertex AI wraps these models in an entire platform, offering tools for serving and scaling your applications effectively. Let’s dive into some of the key Claude models and their strengths.

• Claude Opus 4 and the even newer Claude Opus 4.1: These are Anthropic’s most powerful models to date. They excel at complex, long-running tasks and agent workflows, with sustained performance in areas
  • Advanced coding work
  • Autonomous AI agents
  • Agentic search and research
  • Complex problem-solving
  • Precise context management
• Claude Sonnet 4: This is Anthropic’s mid-size model, offering a balance of performance with speed and cost, making it ideal for high-volume use cases. It surpasses its predecessor, Claude Sonnet 3.7, in coding and reasoning, responding more precisely to steering. Use cases include
  • Every day, coding tasks like code reviews and bug fixes
  • AI assistants
  • Efficient research
  • Large-scale content generation and analysis
• Claude 3.5 Sonnet: Optimized for specific needs, excelling in real-world software engineering such as
  • Agentic tasks
  • Document Q&A
  • Visual data extraction
• Claude 3.5 Haiku: Anthropic’s fastest and most cost-effective model, suitable for use cases where speed and affordability are paramount.
  • Code completions
  • Interactive chatbots
  • Data extraction
  • Real-time content moderation

Optimizing Performance and Cost on Vertex AI

Vertex AI provides developers with crucial tools to manage costs and optimize the performance of Claude models:

• Token Counting: Before sending a large or complex prompt, you can use Vertex AI’s dedicated count-tokens endpoint to determine the number of tokens in a message. This helps predict costs, stay within token limits, and optimize prompts for efficiency, with no cost for using the endpoint itself.
• Prompt Caching: This powerful feature reduces latency and costs for repeated content across multiple requests. When enabled, Vertex AI automatically caches parts of your input, allowing subsequent identical queries to leverage cached results, avoiding redundant computation. This can lead to significant cost reductions, potentially up to 90% savings for frequently used prompts from the second use onwards. The cache lasts for five minutes and refreshes with each access.
• Batch Predictions: For tasks prioritizing throughput over immediate responses, such as generating unit tests or analyzing extensive logs, Vertex AI allows you to batch many prompts into a single request, which is highly efficient for bulk processing. Input datasets can be prepared as BigQuery tables or JSONL files in Cloud Storage.
• Provisioned Throughput: Beyond pay-as-you-go pricing, you can reserve dedicated capacity and prioritized processing for critical production workloads at a fixed fee.
• Request-Response Logging: For monitoring usage and troubleshooting, Vertex AI supports 30-day request-response logging of your prompt and completion activity.
• Built-in Security, Compliance, and Data Governance: Leveraging Google Cloud’s robust security framework, Vertex AI ensures your Claude-powered applications adhere to enterprise standards.

Building Intelligent Agents: Claude, ADK, and Agent Engine

One of the most exciting applications of Claude models on Vertex AI is in building AI agents. Google’s Agent Development Kit (ADK) is designed for flexibility, allowing seamless integration of various large language models, including Anthropic’s Claude, into your agents.

Agent Development Kit (ADK)

This is an open-source, code-first toolkit for developing, evaluating, and deploying AI agents. It provides core concepts like;

• LlmAgent: The agent’s “brain,” defined with a model (like Claude), instructions, and tools.
• Tools: The agent’s skills, typically Python functions, with ADK simplifying the schema generation for the LLM

💡 To learn more about Agent Development Kit (ADK), check our previous blog post here.

Model Context Protocol (MCP)

A powerful way to extend your ADK agent’s capabilities is by leveraging the Model Context Protocol (MCP). MCP is an open standard that standardizes how LLMs communicate with external applications, data sources, and tools, acting as a universal connector. An ADK agent can function as an MCP client, consuming tools exposed by external MCP servers to interact with a wide array of systems, from local file systems to remote APIs. The MCPToolset class in ADK bridges to MCP servers, handling connections and tool discovery.

Vertex AI Agent Engine

Once you have developed your agent with ADK and Claude, Vertex AI Agent Engine becomes crucial for deploying, managing, and scaling it in production. Instead of manually configuring infrastructure or Docker containers, Agent Engine allows you to simply call agent_engines.create() with the Vertex AI SDK.

• Fully Managed Runtime: Agent Engine handles deployment, auto-scaling, regional expansion, and container vulnerabilities.
• Integrated Observability: It comes with built-in integrations for Cloud Trace, Cloud Logging, and Cloud Monitoring, providing insights into agent behavior, “thought processes,” tool interactions, and performance metrics like latency and token usage. This is critical for debugging, optimizing, and iterating on agents with confidence.
• Session Management and Memory Bank: To overcome the “goldfish challenge” where LLMs might lose context across turns in a conversation, Agent Engine offers managed sessions to store conversation histories and maintain conversational flow. Additionally, the Memory Bank persistently stores user memories, automatically extracting and consolidating user facts using LLMs, and providing simple or similarity-based retrieval for relevant context.
• Evaluation and Example Store: Agent Engine integrates with Vertex AI Gen AI Evaluation services for measuring agent performance and offers an Example Store to provide a few-shot examples, improving the quality and consistency of responses.

Agent 2 Agent Protocol (A2A)

A2A is a new standard designed to let agents communicate and collaborate seamlessly, regardless of the framework, runtime, or vendor. Think of it as an interoperability layer for AI agents, just like APIs allow different applications to talk to each other, A2A allows different agents to share messages, delegate tasks, and build workflows together. For teams building enterprise AI solutions, this means you are no longer locked into a single ecosystem. An ADK-powered Claude agent could, for example, collaborate with another vendor’s specialized agent to handle tasks like scheduling, market analysis, or document processing. By using A2A, you unlock a multi-agent architecture where agents can talk, exchange results, and chain capabilities across organizations and platforms.

💡 To learn more about Agent 2 Agent Protocol (A2A), check our previous blog post here.

AgentSpace

While A2A defines how agents talk to each other, AgentSpace defines how enterprises deliver them to people. AgentSpace is an enterprise-grade platform designed to give employees access to powerful AI agents and enterprise data sources in a secure, managed environment. Instead of AI being siloed in experimental labs or single apps, AgentSpace brings it into the day-to-day workflow, enabling teams to use Claude, Gemini, or domain-specific agents directly for tasks like drafting proposals, analyzing market growth, or summarizing customer feedback. IT and business leaders benefit from centralized governance, compliance, and observability, while employees benefit from AI that feels embedded into their work. In short, AgentSpace is the distribution layer: it operationalizes AI agents across the enterprise, ensuring they are not just built, but actually used at scale.

💡 To learn more about AgentSpace, check our previous blog post here.

🎥 Prefer watching instead of reading? We have created a NotebookLM podcast video with slides and visuals based on this blog.

⭐⭐⭐

The integration of Anthropic’s Claude models with Vertex AI’s enterprise platform creates a powerful foundation for building the next generation of intelligent agents. From streamlining workflows with Claude Code, to building multi-agent systems with ADK and MCP, to scaling securely with Agent Engine, A2A, and AgentSpace, Google Cloud provides everything you need to move from experimentation to enterprise adoption.

Now is the time to transform how your teams work with AI. Contact us today to start building and deploying AI agents on Vertex AI, bringing automation, intelligence, and scalability directly into your business.