According to its own statements, Google Cloud is the first major cloud provider to offer all three of the currently leading model families via a single platform: its own Gemini family, Claude from Anthropic and now also the GPT models from OpenAI. At launch, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna are available as a private preview for enterprise customers, i.e. in restricted access ahead of general availability.

What sounds like a pure portfolio expansion changes one thing above all for companies: the architecture decision at the start of an AI project. Until now, choosing the model provider often coincided with choosing the cloud environment. Anyone who wanted to use several model families generally had to manage several access points, map different billing models and go through separate security and data protection reviews.

Why model diversity matters in practice

In real projects there is rarely one model that solves every task equally well. A language model that reliably summarises long technical texts is not automatically the best choice for structured data extraction, for code generation or for a dialogue assistant in customer service. Add to that non-functional criteria: response times, cost per request, availability in certain regions and the question of how well a model fits into existing approval processes.

When several model families are reachable via the same platform, such trade-offs can be made per use case instead of once across the board for the entire company. A product configurator can then use a different model than the internal knowledge search, without a second infrastructure being created for it.

Dependence on the provider becomes more calculable

The second effect concerns what IT calls lock-in: the technical and contractual tie to a single provider that makes a later switch expensive. This tie does not disappear through a broader model offering; it merely shifts. Dependence on the individual model provider decreases, while dependence on the platform through which everything runs tends to increase.

For project planning, this is still progress. A model change within the same platform ideally affects only the application's calling layer and not the surrounding infrastructure of authentication, logging, cost control and data storage. It is precisely this environment that causes the larger part of the effort in practice.

What can be derived from this for the architecture

Regardless of which platform a company uses, a setup that keeps the specific model interchangeable is advisable. A few principles have proven themselves:

  • Abstraction layer: The application does not talk directly to a model endpoint, but to an internal interface behind which the model can be swapped.
  • Prompts as configuration: Instructions to the model should be versioned and managed outside the program code so that they can be adapted per model.
  • Measurable quality: A fixed set of test cases with expected results makes the comparison of two models verifiable instead of leaving it to gut feeling.
  • Cost and runtime data: Without logging token consumption and response times, any profitability calculation remains an estimate.
  • Fallback level: For productive processes, it should be defined what happens if a model is unavailable or delivers unusable responses.

Private preview does not mean production-ready

The fact that the GPT models are initially available only as a private preview for enterprise customers deserves attention. In this phase, availability, terms and scope of functions may still change. For evaluations, prototypes and internal comparison tests, this is a sensible framework. For business-critical processes with commitments to customers, however, it is advisable to wait until clear operating and contractual conditions are available.

A staged approach is therefore sensible: first a narrowly defined use case with manageable risk, then a robust comparison of several models based on your own data, and only then the decision on productive use. Anyone who follows this sequence avoids building an architecture around a model that has already been superseded in six months.

Data protection and governance remain separate tasks

Shared platform access simplifies the technical connection but does not replace legal review. Which data may be transferred to which model, in which regions processing takes place and how long inputs are stored must still be clarified per use case. Especially with personal or contract data, this review should take place before the first prototype and not after.

An internal approval list that records which models are permitted for which data classes is helpful. This creates clarity for development teams and prevents decisions from being improvised on a case-by-case basis.

Classification for web and software projects

For web and software projects, the central question shifts from "Which provider do we choose?" to "How do we keep the model choice open?". Projects should be cut in such a way that a model change remains a configuration change and does not trigger a rebuild of the application.

Anyone who sets up the abstraction layer, test cases and logging cleanly today can evaluate new models later with little effort, instead of experiencing every market movement as a project risk.

Sources