The era of the single-model enterprise AI stack is over.

While a massive, reasoning-heavy AI model might be perfect for resolving complex contractual ambiguities, using it to extract bulk invoice data is like driving a race car in rush-hour traffic—expensive and inefficient. Conversely, smaller, highly optimized models can excel at low-latency tasks but may lack the depth required for some advanced agentic workflows.

Enterprise AI increasingly requires a multi-model portfolio, and Google Cloud has just fundamentally shifted the landscape to support this. With the addition of OpenAI models to its Model Garden, Google Cloud is now the only hyperscaler offering access to all three leading closed-source LLMs, alongside more than 200 first-party, third-party, and open models on a unified enterprise AI platform.

By bringing OpenAI alongside Gemini, Anthropic Claude, xAI, Meta, Mistral AI, AI21 Labs, and a broad selection of open-model alternatives, Google Cloud gives businesses more ways to evaluate and deploy AI capabilities for different workloads within a common enterprise cloud environment. The available models, routing options, and security controls depend on the selected model and deployment method.

For CIOs, CTOs, and enterprise architects, model choice is no longer just a technical preference; it is a foundational architectural decision driving application design, security, procurement, and long-term agility.

Beyond Gemini: A Broader Managed Model Ecosystem

Gemini remains a central part of Google Cloud’s AI portfolio, but it’s not the only option. Through Model Garden and the Gemini platform, organizations can access models from external providers via managed APIs and explore self-deployment options for supported models. Managed endpoints let teams use selected models without provisioning or operating their own model-serving infrastructure.

Crucially, this approach can simplify enterprise procurement. OpenAI models purchased through Google Cloud Marketplace appear on a single Google Cloud invoice, draw down existing cloud commitments, and count toward Flexible Savings Plans (FSP). This can reduce billing fragmentation and simplify cost management across AI workloads. Organizations should verify the applicable commitment and pricing terms for each model.

Here is a selection of third-party models available through Google’s managed APIs, including offerings in preview. Availability and access methods vary by model.

Provider	Available Model Families & Examples	Enterprise Applications
OpenAI	GPT-6 Astra, GPT-6.1 Sol, and GPT-6 Luna, currently in private preview	Advanced coding, computer use, scientific research, professional work, and high-volume everyday tasks.
Anthropic	Claude Haiku 5.5, Claude Sonnet 5.5, Claude Opus 5.5, and other supported Claude versions	Complex reasoning, coding, agentic workflows, document analysis, and professional knowledge work.
xAI	Grok 4.3, Grok 4.6, Grok 4.7, and Grok 4.20, where supported	Coding, reasoning, tool-using agents, and other knowledge-intensive tasks.
Meta	Muse Spark 1.3, a managed partner model, alongside the separate Llama open-weight family	Reasoning, multimodal applications, and agentic workflows, depending on the model’s supported capabilities.
Mistral AI	Mistral Medium 3, Mistral Small 3.1, Mistral OCR 25.05, and Codestral 2, where supported	General language tasks, document OCR, summarization, programming, and code completion.
AI21 Labs	Jamba 1.5 Large and Jamba 1.5 Mini, listed in preview	High-throughput text generation, document processing, and workloads where response quality, speed, and cost need to be balanced.
DeepSeek	DeepSeek-V4.1-Flash, available as a managed API in preview	Agentic workflows, coding, and multimodal tasks involving text and images

Open-Weight Models: Greater Control Over Inference and Deployment

For organizations that need greater control over inference infrastructure, deployment architecture, and model customization, the ecosystem extends far beyond managed APIs.

Google Cloud’s Model Garden includes a broad selection of open-model options, including DeepSeek, Meta’s Llama family, Mistral and Mixtral, Kimi, MiniMax, Qwen, and models from Z.AI. Supported deployment paths and serving frameworks, such as vLLM, give engineering teams options for running suitable models on Google Cloud infrastructure, including within a VPC when the chosen architecture supports it. Google’s catalog also includes managed partner models, such as AI21 Labs’ Jamba 1.5 Large and Mini, which are listed separately from its open-model offerings.

This approach presents a strategic trade-off: organizations assume greater responsibility for infrastructure, including capacity planning, patching, scaling, and model lifecycle management, in exchange for greater control over model serving, customization, and supported tuning options. The resulting privacy and compliance posture depends on how the model is deployed and governed.

Specialized Models for Specialized Architectures

Not every AI problem requires a general-purpose conversational model. The most efficient enterprise architectures combine general reasoning with specialized, purpose-built components.

Gemini platform supports specialized partner models and deployment options for selected offerings. Depending on the model, organizations may be able to license and self-deploy it within their own cloud environment or VPC, subject to the applicable licensing terms and supported configurations.

Provider	Model	Primary Enterprise Purpose
Contextual AI	Reranker	Improving relevance and accuracy in Retrieval-Augmented Generation (RAG).
CAMB.AI	MARS7	Multilingual text-to-speech, voice cloning, and expressive speech generation.
Virtue AI	VirtueGuard	AI safety guardrails, real-time content security, and strict policy enforcement.
Qodo	Embed-1	Specialized code embeddings for vast enterprise codebase retrieval.
Mistral AI	Codestral	Specialized code generation and code completion.
WRITER	Palmyra X4	Structured output generation, rigorous tool calling, and RAG synthesis.
AI21 Labs	Jamba Large 1.6	High-quality language generation and efficient private enterprise deployment.
CSM	Cube	Generating 3D digital assets from 2D reference images.

Consider an enterprise RAG pipeline for software engineering: an architecture might use Qodo Embed-1 to create embeddings for source code and related technical content, Contextual AI’s Reranker to improve the relevance of retrieved results, and Anthropic’s Claude Sonnet to synthesize the final answer.

Model Availability and Documentation

Google Cloud’s AI model catalog evolves frequently. Model availability, model versions, preview and general availability status, regional access, deployment options, and compatibility with features such as Smart Routing may change over time. The models listed in this article reflect the information available at the time of writing and may not be available across all Google Cloud services, deployment methods, or regions.

Before selecting a model or designing a production deployment, consult Google’s official Partner Models documentation and Model Garden documentation for the latest availability, supported configurations, and limitations.

Dynamic Execution: Smart Routing and Orchestration

The ultimate value of a multi-model strategy isn’t just having access to more models; it is the ability to route tasks dynamically without hard-coding a specific provider into your application logic.

Google’s Smart Routing automatically triages supported enterprise workloads, directing them to the model that best balances performance and cost. An enterprise application can combine this capability with retrieval systems and business logic to coordinate a multistep customer service workflow.

For example:

• Step 1: A fast, low-cost model classifies the intent of incoming tickets.
• Step 2: A specialized retrieval model pulls relevant policy data.
• Step 3: A capable reasoning model, selected from the models supported by the relevant orchestration workflow, handles complex edge cases and drafts the final response.

By abstracting the model layer from the business logic, organizations can improve operational consistency, mitigate vendor lock-in, and adapt their stack more readily as model capabilities and pricing evolve.

Note: Google currently describes model orchestration across the Gemini family and Anthropic Claude today, with other leading private and open models planned for the future. Availability through Model Garden doesn’t automatically mean that a model is supported by Smart Routing. Confirm the eligible models and capabilities for the specific product and deployment.

Enterprise Data Perimeters: Security Must Follow the Model

Greater model choice demands rigorous, unified governance. Running models across multiple providers introduces additional questions about data handling, access control, residency, and auditing.

Google Cloud describes specific enterprise security controls for model access through its platform. For Gemini, managed partner APIs, and self-deployed open models, organizations should verify the protections and responsibilities that apply to their chosen service, endpoint, and deployment architecture.

Key governance features include:

• Data protection and model training: Prompts, context, and outputs for the OpenAI offering remain within the designated data perimeter and aren’t used to train foundation models. Zero Data Retention (ZDR) options are announced as coming soon for this offering.
• VPC Service Controls: Supported configurations can use VPC Service Controls to help enforce data perimeters and reduce data-exfiltration risks. Organizations should verify the network controls available for each model, endpoint, and deployment architecture.
• Identity and auditability: Authenticated interactions, identity and access controls, and unified audit logging can help IT teams establish centralized visibility over supported model and agent activity, investigate issues, and support compliance requirements.

Build Your AI Strategy Around Your Business

Google Cloud’s expanding model ecosystem empowers enterprise teams to align AI capability, cost, and deployment requirements directly with business outcomes. But moving from a single-model pilot to a secure, multi-model production environment requires rigorous architectural planning.

To succeed, organizations must build an abstraction layer for their APIs, define strict evaluation criteria, run comprehensive latency and cost tests, and establish robust data governance before scaling.

Ready to architect a future-proof AI ecosystem? Kartaca helps enterprises turn Google Cloud’s vast AI capabilities into production-ready business solutions. From multi-model evaluation and API abstraction to secure VPC deployment, integration, and operational governance, we bridge the gap between AI potential and enterprise reality.

Contact us today to identify the right models, routing strategies, and deployment architectures for your workloads.