Microsoft has integrated Anthropic's Claude models into Microsoft 365 Copilot. This is not a minor upgrade: Claude now powers features across Researcher, Excel Agent Mode, Word and PowerPoint agents, Copilot Studio, and the full Copilot Chat experience.
For organizations using Copilot, this raises practical questions: which Anthropic models are included, which features use Claude, what does it cost, what do admins need to do, and where is your data processed? Here is what you need to know.
Which Anthropic models are in Copilot?
Microsoft offers the current Claude generation from Anthropic in Copilot alongside the OpenAI models. You do not need a separate Anthropic license: the models are covered by your Microsoft 365 Copilot license. New Claude versions are added as Microsoft releases them; the Microsoft 365 admin center shows which versions are currently active.
| Claude family | Best at | Where in Copilot |
|---|---|---|
| Claude Sonnet | Everyday work: writing, summarizing, analysis. Balance of speed and quality | Researcher, Copilot Chat, Agent Mode in Excel, Office agents, Copilot Studio |
| Claude Opus | Complex reasoning, long documents, code | Researcher (critique step), Copilot Studio |
| Claude Haiku | Fast, low-cost agent steps such as classification and routing | Copilot Studio |
| Anthropic Preview models with Data Retention (Claude Fable, Mythos) | Latest preview models | Separate opt-in per tenant, see below |
In short: Sonnet is the default choice, Opus when the answer really has to be better, Haiku when speed and cost matter most.
Where Claude is already active
Claude models are available across a growing set of Microsoft 365 Copilot features:
- Researcher: Claude works alongside GPT in a "Critique" workflow. GPT drafts a response to a research query, then Claude reviews it for accuracy, completeness and citation quality before delivering the final result. This multi-model approach led to a 13.8% improvement on the DRACO benchmark, outperforming standalone deep-research tools from OpenAI, Google, Perplexity and Anthropic.
- Copilot Chat: Claude is available across the full Copilot Chat experience (web, desktop and mobile). A UI indicator shows when Claude is in use.
- Agent Mode in Excel: Users can select Claude as the model for data analysis tasks.
- Word, Excel and PowerPoint agents: Claude powers agent capabilities in Office apps.
- Copilot Studio and Power Platform: Creators can select Claude during agent creation.
Anthropic in Copilot Studio
In Copilot Studio you choose which model an agent uses, both for the agent's orchestration and per prompt or tool. Next to the OpenAI models you find the Anthropic models, provided the Anthropic toggle is enabled in your tenant (see below). That enables combinations: Claude Haiku for fast classification and routing steps, Claude Sonnet or Opus for the substantive answers, and GPT where it performs better.
Cost does not depend on the model you pick: agents in Copilot Studio are billed through Copilot Credits or messages, regardless of the model provider. Note that Anthropic processing here also takes place in the US, under the same subprocessor terms as in Microsoft 365 Copilot. If you need an agent that stays entirely within your own Azure environment, that is possible with a custom build on Azure; see our AI Cloud solutions.
Claude or GPT: when to pick which model?
For most users the model choice is invisible: Copilot picks for you, and in Researcher GPT and Claude already work together. Where you can choose (Agent Mode in Excel, Copilot Studio, the model picker in Copilot Chat), this is a workable rule of thumb:
- Claude for long documents, structured analysis, code and tasks where careful reasoning matters more than speed.
- GPT for quick, short interactions and tasks deep in Microsoft Graph data, which the default Copilot experience is tuned for.
- Both for important output: let one model draft and the other review, the way Researcher does automatically.
Admin controls: what you need to configure
Anthropic is onboarded as a Microsoft subprocessor since January 7, 2026. This means Claude operates under Microsoft's Product Terms and Data Protection Addendum (DPA), not under a separate Anthropic agreement.
Default state depends on your region
| Region | Default state | Action needed |
|---|---|---|
| Most commercial cloud (US, APAC, etc.) | On by default | Review and disable if needed |
| EU / EFTA / UK | Off by default | Opt in if you want Claude features |
| Government clouds (GCC, GCC High, DoD) | Not available | No toggle present |
How to enable or disable
- Go to the Microsoft 365 admin center → Copilot → Settings
- On the Data access page, select AI providers operating as Microsoft subprocessors
- Toggle Enable Anthropic as a Microsoft subprocessor on or off
You need the Global Administrator role to change this setting. Once disabled, users lose access to all Claude-powered features.
Per-user and per-group controls
Admins can assign third-party model providers to specific users or Entra ID security groups. This means you can enable Claude for your data analysis team while keeping it off for the rest of the organization. The setting works at the provider level (not per model) and supports up to 999 groups with nested group membership.
EU vs. US: where is your data processed?
This is the most important consideration for European organizations:
Anthropic processing happens in the United States. Claude models deployed in Microsoft 365 Copilot are currently excluded from the EU Data Boundary and in-country processing commitments.
What this means in practice:
- EU/EFTA/UK tenants: Anthropic is off by default. If you enable it, you are explicitly opting in to data processing outside the EU Data Boundary.
- US and other regions: Anthropic is on by default. Data is processed in the US under Microsoft's DPA.
- All regions: Anthropic does not train on your data. Prompts and outputs are not stored persistently by Anthropic.
Compliance framework
When Anthropic operates as a Microsoft subprocessor, the following apply:
- Microsoft Product Terms and DPA: govern all data processing
- Microsoft Enterprise Data Protection: covers Copilot interactions using Claude
- Microsoft Customer Copyright Commitment: applies to Claude output in M365 Copilot and Copilot Studio
- Microsoft audit controls: Purview, eDiscovery, and compliance tools work as usual
The key distinction: your data flows through Microsoft infrastructure with Microsoft oversight. Anthropic processes it under Microsoft's contractual safeguards, not as an independent service. For how Enterprise Data Protection for Copilot works in detail, read our article on Copilot and enterprise data protection.
Legacy toggle deprecation
If you previously enabled Anthropic under the old toggle (which used Anthropic's own commercial terms and DPA), that option has been deprecated as of January 7, 2026. EU/EFTA/UK tenants that previously opted in need to re-enable under the new subprocessor toggle. The old agreement no longer applies.
Exception: Anthropic Preview models with Data Retention
Most Claude models run with Anthropic as a Microsoft subprocessor, under Microsoft's contracts. But there is one category to watch: Anthropic Preview models with Data Retention (such as Claude Fable and Mythos).
For these preview models, Anthropic acts as an independent processor, not a Microsoft subprocessor. They are governed by Anthropic's own Commercial Terms and Data Protection Addendum, and Anthropic retains inputs and outputs (typically up to 30 days, and longer for trust-and-safety review). Anthropic does not use retained data for training without your explicit permission.
These models are off by default for every tenant, even where standard Anthropic models are on, and require a separate, explicit admin opt-in in the Microsoft 365 admin center. If your compliance posture matters, treat preview-with-retention models as a deliberate, separate decision.
Recommendations for IT admins
- Check your current toggle state: Go to Copilot → Settings → Data access and verify whether Anthropic is enabled or disabled for your tenant
- Assess EU Data Boundary requirements: If your organization requires all data processing within the EU, keep Anthropic disabled until Microsoft offers EU-based Claude processing
- Communicate to users: Claude-powered features show a UI indicator, but users should understand what this means
- Use per-user controls for a targeted rollout: If you want Claude for specific teams only, assign the provider to those users or Entra ID security groups in the admin center
- Update your AI governance documentation: Document your decision to enable or disable Anthropic, and the rationale behind it
Get help
Need help reviewing your Copilot settings or assessing the impact of Anthropic models on your compliance posture? Contact Universal Cloud. We help organizations configure Microsoft 365 Copilot responsibly and get the most out of AI.
