7 Min. Read Time
Microsoft has opened up the model question with Copilot Wave 3. In Microsoft 365 Copilot, Anthropic’s Claude models now run alongside those from OpenAI, within the same product, sometimes in the same workflow. For cloud and platform teams, this isn’t a comfort feature but an architectural decision. It affects data flow, latency, costs, and the question of how many AI providers a company ultimately depends on.
Key Takeaways
- Claude runs alongside GPT: With Wave 3, Anthropic’s Claude models are available in the Copilot main chat, in addition to OpenAI models. Copilot automatically selects the suitable model for each task.
- DACH gets a special rule: In the EU, the UK, and the EFTA, an admin opt-in is still required for Anthropic models. Without this step, the stack remains with OpenAI.
- A second provider means a second subprocessor: Anthropic comes in as an additional subprocessor in data processing. This means less vendor lock-in but more contract processing that an audit must track.
Related:Copilot Studio Becomes the Agent Control Center / AWS Bedrock AgentCore Restricts Agent Strategies
What Model Choice Means Technically
Model Choice is the feature that allows Microsoft 365 Copilot to no longer be tied to a single language model. Until 2025, Copilot ran exclusively on OpenAI models. With Wave 3, Anthropic’s Claude models have been added, in the main chat via the Frontier program and in Copilot Studio as a selectable alternative for self-built agents.
The mode that affects cloud teams the most is automatic routing. According to Microsoft, Copilot selects the model that best fits each task. Reasoning-intensive tasks can go to Claude, while others go to an OpenAI model. In Copilot Studio, OpenAI remains the default for new agents; selecting Claude is a conscious decision by the creator.
It’s essential to understand: Model Choice isn’t a neutral switch. Once routing is active, the same workflow in a session can run through two providers. This is the actual innovation. It has technical consequences that go beyond pure model quality.
The Data Flow and a New Subprocessor
The first consequence concerns data flow. If a request goes to a Claude model, Anthropic becomes a subprocessor in the processing chain of Microsoft Online Services. Microsoft lists Anthropic accordingly in its subprocessor documentation. For a company whose data protection audit checks the contract processing chain, this is a new entry that must be documented and evaluated.
This is also where the most important DACH peculiarity lies. Anthropic models have been standardly available in Copilot Studio in most regions since January 1, 2026. In the EU, the UK, and the EFTA, this does not apply automatically. An explicit admin opt-in is required there before the Claude models are used. If you operate Copilot in a German company and don’t change anything, you’ll stay on the OpenAI path for now.
This isn’t a disadvantage but a decision situation. The opt-in is the point where the data protection and IT functions should look together. Which data classes does Copilot process? Does the company want to put this data into a second provider’s infrastructure? Is the contract processing cleanly regulated for this? These questions belong before the opt-in, not after.
Routing, Latency, and the Cost Question
The second aspect is operational in nature. A stack with two model providers behaves differently than one with a single provider. The following overview shows where the two model families stand in the Copilot context.
| Aspect | OpenAI Models | Claude Models |
|---|---|---|
| Role in the Stack | Default for Copilot and new agents | Selectable alternative, switchable in routing |
| Strength according to Microsoft | Broad all-round suitability, creative content | Reasoning-intensive tasks |
| DACH Availability | Available | EU, UK, and EFTA only with admin opt-in |
| Data Processing | OpenAI as sub-processor | Anthropic as additional sub-processor |
Source: Microsoft 365 Blog and Copilot Studio Blog on model selection, as of May 2026.
When it comes to latency, honesty is in order. Microsoft does not publish reliable response time comparisons between models in the Copilot context. If latency is a criterion, it must be measured in your own tenant, with real prompts from your daily work. General statements that one model is generally faster do not apply in a routing setup, as the task and not the model determines the path.
The cost question is different in the Copilot context than with direct model API access. Microsoft 365 Copilot is licensed per user. Changing the model in the main chat does not directly affect the license calculation. In Copilot Studio, however, usage-based components may apply, and the cost side needs to be examined more closely. What changes in both cases is the predictability. In a single-model stack, the behavior is more consistent. In a routing stack, the result depends on which model Copilot chooses. And this can shift with Microsoft updates.
Vendor dependency: less or more
The strongest argument for Model Choice is the reduced dependency on a single AI provider. The argument is valid, but it’s shorter than it sounds. Anyone who lists both sides soberly makes a better decision.
What Model Choice alleviates
- No total failure if a single model provider encounters problems
- Tasks can be assigned to the stronger model
- The negotiating position with a single provider improves
- Strategic freedom of choice if the model quality shifts
What Model Choice shifts
- The dependency on Microsoft as a platform remains unchanged
- A second sub-processor extends the order processing chain
- The routing is a black box controlled by Microsoft, not the customer
- The behavior becomes less predictable because the path varies per task
The crucial point is listed at the top of the right column. Model Choice reduces dependency on OpenAI or Anthropic, not on Microsoft. Copilot remains the platform, routing remains in Microsoft’s hands, and the license remains the same. Anyone who sells Multi-Model as a sovereignty gain confuses the model level with the platform level.
What cloud teams need to clarify now
Model Choice is a good development, but not one that happens automatically. Three points need to be on the table before the opt-in is set.
Firstly, the data question. Before the EU opt-in, the data protection function clarifies whether the data classes processed by Copilot are allowed in a second provider infrastructure. The order processing with Anthropic as a sub-processor is documented, not taken silently.
Secondly, measurement. Anyone expecting a quality or speed gain from Model Choice measures it in their own tenant with real prompts. Without this measurement, model choice remains a gut feeling that Microsoft updates can shift at any time.
Thirdly, expectations. Model Choice is a tool against provider cluster risk at the model level. It is not an exit strategy from the Microsoft ecosystem. This distinction belongs in every architecture discussion; otherwise, a false sense of security arises.
The sober classification: Claude alongside GPT in the same stack is progress because it creates real choice where there was none before. However, the progress ends at the platform boundary. Anyone who understands this uses Model Choice for what it can do. They plan the platform question further as what it is.
Frequently Asked Questions
What is Model Choice in Microsoft 365 Copilot?
Model Choice is the feature that frees Copilot from being tied to a single language model. With Wave 3, Anthropic’s Claude models are now available alongside OpenAI models. Copilot automatically selects the most suitable model for each task, while in Copilot Studio, the model can also be set manually per agent.
Which models are available?
In Copilot Studio, an OpenAI model remains the default for new agents. Additionally, Anthropic’s Claude models can be selected—such as Claude Sonnet and Claude Opus—in the agent builder. In the main chat, Claude is available through the Frontier program.
Are the Claude models available in Germany?
Not automatically. In the EU, the UK, and the EFTA, explicit admin opt-in is required before Anthropic models can be used. In most other regions, they have been enabled by default since early January 2026. Without opt-in, a German Copilot tenant remains restricted to OpenAI models.
What does Model Choice mean for data privacy?
Once a Claude model is used, Anthropic becomes an additional subprocessor in Microsoft Online Services’ data processing chain. This is a new entry that must be documented and assessed during a data protection audit. Therefore, the opt-in decision should be made jointly by IT and data protection teams.
Does Model Choice reduce dependence on Microsoft?
No. Model Choice reduces dependency on a single model provider like OpenAI or Anthropic. The platform remains Microsoft Copilot, routing is controlled by Microsoft, and the license stays the same. Multi-model support is an advantage at the model level—not an exit strategy from the ecosystem.
More from the MBF Media Network
