Microsoft on September 24, 2025 opened Microsoft 365 Copilot to a new family of large language models: Anthropic’s Claude Sonnet 4 and Claude Opus 4.1. The change ends the de facto OpenAI exclusivity inside Copilot’s productivity surface and marks the company’s most deliberate step yet toward multi-model orchestration—letting organizations route specific workloads to the model best suited for the job, all while keeping OpenAI and Microsoft’s own models in the mix.

What Actually Changed

The update surfaces Anthropic models in two Copilot experiences immediately:

  • Researcher agent — Users can now choose Claude Opus 4.1 as an alternative reasoning engine for deep, multi-step research and report generation. A “Try Claude” button appears in the Copilot interface, letting workers switch between OpenAI and Anthropic models on the fly.
  • Copilot Studio — The low-code agent builder now lists both Claude Sonnet 4 and Claude Opus 4.1 as selectable model options for custom agents. Developers can mix models within a single agent, combining a Sonnet-powered task with an OpenAI-based step, for example.

These additions are not automatically available to every tenant. Microsoft is rolling out model choice through the Frontier early-access program and gradually to all licensed organizations. Critically, tenant administrators must enable Anthropic models before they appear for end users—an intentional gate that puts control squarely in IT’s hands.

Physically, the Anthropic models are not running on Azure. They remain hosted on Amazon Web Services, Anthropic’s primary cloud home, and are accessed via the standard Anthropic API. Microsoft confirmed that Copilot calls to Claude will traverse public cloud networks, introducing cross-cloud inference and billing that differ from the all-Azure pipeline many enterprises have come to expect.

What It Means for You

For the everyday Word, Excel, or Teams user, today’s change is subtle. Copilot still looks and feels the same. But behind the scenes, some requests may now be handled by a different kind of brain—one that excels at structured data transformation, precise spreadsheet logic, or long-document synthesis. The practical payoff: tasks that chugged before might feel zippier, and outputs in specific domains could be more accurate.

For IT professionals, security officers, and procurement leads, the implications are material. Copilot is no longer a single-vendor dependency. That shift brings several new layers to manage:

Model selection becomes a policy vector. You’ll need to decide which job functions are permitted to use Anthropic models, and for which data classes. Not every request belongs on a third-party cloud. Sensitive PII, PHI, or regulated records should be screened and potentially blocked from cross-cloud inference.

Cross-cloud data flows are real. When a user prompts Copilot and the orchestrator routes that prompt to Claude on AWS, the data leaves your Azure tenant. This triggers data residency, egress, and compliance questions that didn’t exist when all inference stayed inside Microsoft’s domain. Check your industry regulations and update data loss prevention (DLP) rules accordingly.

Billing and contracts need a fresh look. Microsoft says end-user Copilot pricing won’t change immediately, but the underlying money flow—Microsoft paying Anthropic (and indirectly AWS) for API calls—could eventually influence your enterprise agreement. Ask your Microsoft account team for clarity on cross-cloud service-level agreements, data processing addenda, and indemnities related to Anthropic-hosted calls.

Audit trails just got more important. If a regulator or internal investigation asks why Copilot generated a specific answer, you must know which model served the request. Ensure Copilot telemetry captures model provenance and that logs are retained per your retention policies.

How We Got Here

Microsoft’s Copilot journey began with a deep, multi-billion-dollar bet on OpenAI. For two years, GPT-4 and its successors powered nearly every Copilot feature. That partnership accelerated generative AI’s arrival in Office, but it also concentrated risk.

By mid-2025, three forces pushed Microsoft toward diversification:

  • Scale economics: Running billions of inferences a day on a single family of frontier models is costly. Smaller, task-optimized models like Claude Sonnet 4 can deliver acceptable quality at a fraction of the compute cost.
  • Task fit: Benchmarks and internal testing repeatedly showed that different models shine on different workloads. Anthropic’s Claude family, for instance, has earned a reputation for strong structured-output tasks—exactly the kind of heavy lifting found in Excel automation or slide generation.
  • Vendor leverage: Reliance on one external AI provider gave that provider outsized negotiating power. By bringing in Anthropic, Microsoft spreads its bets and gains commercial flexibility, all while continuing to invest in its own MAI models.

The September 24 announcement didn’t come out of nowhere. A week earlier, Microsoft began favoring Claude Sonnet 4 inside Visual Studio Code’s GitHub Copilot, automatically selecting it for many coding tasks based on performance. Reports that Microsoft found Anthropic models outperformed OpenAI’s in Excel and PowerPoint testing also suggested Copilot would expand model choice. Now it has.

What to Do Now

If you’re an IT decision maker, don’t wait for a broader rollout to start preparing. Here’s a concrete checklist:

  1. Locate the admin controls. In your Microsoft 365 admin center, find the Copilot settings section. Identify where Anthropic models can be toggled and who has permission to flip that switch. Lock it down to a small, trusted group.
  2. Run a pilot with non-sensitive workloads. Pick a handful of teams—marketing for slide decks, sales for contract analysis—and test Claude-driven Researcher or Copilot Studio agents. Compare outputs against baseline OpenAI results. Measure both quality and latency.
  3. Update DLP and classification policies. Create rules that flag or block prompts containing sensitive information from routing to Anthropic models. Categories like health data, financial records, and personally identifiable information should trigger protective action.
  4. Review contracts and terms. Request explicit documentation from Microsoft about how Anthropic API calls are handled contractually, where data is processed, and what happens in a breach scenario. If your compliance framework demands data sovereignty, confirm that cross-cloud inference doesn’t violate those rules.
  5. Turn on telemetry and logging. Ensure Copilot usage logs capture the model identifier for every request. This data will be essential for audit, troubleshooting, and cost analysis later.
  6. Communicate with users. Let employees know that Copilot may now use multiple AI models. Provide guidelines on what to expect and how to provide feedback—especially if they notice a change in output quality.

Outlook

Microsoft’s move signals the beginning of an orchestration era. Expect Anthropic support to spread beyond Researcher and Copilot Studio. The company has already hinted at bringing Claude into Excel, PowerPoint, and other high-value Copilot experiences where structured outputs are king. Meanwhile, the orchestration layer itself will grow more sophisticated, automatically routing tasks to the optimal model based on cost, latency, and capability.

For enterprises, the multi-model future isn’t a distant concept—it’s a governance challenge arriving today. Those that treat model choice with the same rigor they apply to data access, identity management, and endpoint security will extract the most value. Start the pilot, lock down the policies, and prepare your organization for a Copilot that thinks with more than one brain.