On September 24, 2025, Microsoft began the process of transforming Microsoft 365 Copilot from an OpenAI-exclusive assistant into a multi-model platform. The company announced that licensed Copilot customers—starting with those in its Frontier early-access program—can now opt to use Anthropic’s Claude models (Sonnet 4 and Opus 4.1) inside the Researcher agent and Copilot Studio. The catch? Tenant administrators must flip a switch to enable them. OpenAI’s GPT family still powers the most demanding “frontier” tasks, but for routine productivity work, Claude is now a first-class citizen.

This isn’t a simple feature toggle. It reflects a strategic re-engineering of Copilot toward an orchestration layer that routes each user request to the best-fit model. The change promises cost savings, better performance on certain Office jobs, and a fallback against over-reliance on any one AI lab. But it also hands IT teams new compliance, billing, and observability challenges that have no precedent in the M365 world.

What actually changed

Microsoft’s September 24 announcement introduced two specific Anthropic models into the Copilot ecosystem:

  • Claude Sonnet 4: a mid-sized, cost-efficient model optimized for high-throughput, everyday tasks.
  • Claude Opus 4.1: a larger, more capable model for deeper reasoning and complex work.

These models are initially available inside two Copilot features:

  1. Researcher agent: a tool designed to handle multi-step research and synthesis across documents and data sources.
  2. Copilot Studio: the low-code environment where organizations build custom AI agents, now with the ability to specify a preferred model backend.

The rollout is gated. Only tenants in the Frontier early-access program get immediate access, and even then, a Microsoft 365 administrator must explicitly authorize the Anthropic models before any employee can use them. OpenAI’s models remain the default for all other Copilot interactions, particularly those Microsoft considers “frontier” in complexity.

Beneath the surface, Microsoft has built a dynamic routing engine. When a user invokes Copilot, the orchestration layer analyzes the request—task type, latency tolerance, data sensitivity, cost limits—and selects the optimal model from a pool that now includes OpenAI, Anthropic, and eventually Microsoft’s own in-house MAI models. The user sees the same interface; the model swap happens invisibly.

What it means for you

The impact of this change depends entirely on your role.

For everyday users

If your IT department decides to enable Claude, you might notice subtle differences in Copilot’s output when working in Word, Excel, or PowerPoint. Microsoft’s internal tests, cited by multiple sources, found that Claude Sonnet 4 performs especially well on tasks like spreadsheet automation, slide generation, and visually oriented work. So, the AI may become slightly more fluid or structured in those scenarios. However, since routing is invisible, you probably won’t see a label telling you which model is behind a particular response.

For power users and citizen developers

Copilot Studio gets a meaningful upgrade. When building a custom agent, you can now choose between OpenAI and Anthropic backends, essentially tuning the brain of your agent for specific jobs. If you’ve been prototyping an internal Q&A bot or a document summarizer, you can test which model delivers better accuracy or speed. Just remember: model selection is subject to what your admin has approved.

For IT administrators

This is where the real work begins. You’re now managing a multi-vendor AI supply chain inside your tenant. Key responsibilities include:

  • Explicit opt-in: You must manually enable Anthropic models. If you do nothing, nothing changes.
  • Data residency and compliance: Because Anthropic’s enterprise deployments are primarily hosted on AWS (via Amazon Bedrock), enabling Claude means some user prompts and data will likely travel from Azure to AWS. You’ll need to verify that this cross-cloud flow doesn’t violate your data residency commitments or industry regulations.
  • Billing complexity: Microsoft has not yet detailed how Anthropic inference costs will be passed through. You could see separate line items for third-party model usage, and costs may vary dynamically based on routing decisions. Budget forecasting becomes trickier.
  • Observability: You’ll need to instrument everything—track which model handled which request, measure latency spikes, log error rates, and trace token consumption for chargeback.
  • Output quality and consistency: Different models hallucinate differently and refuse prompts according to their own safety policies. If one department gets a concise answer from Claude while another gets a more elaborate—or refused—response from OpenAI, end-user trust could erode.

For developers and ISVs

If you build apps on top of the Microsoft 365 platform, the orchestration layer introduces new possibilities and testing requirements. An app that calls Copilot APIs may now hit multiple backends unexpectedly. You’ll need to validate that your downstream logic handles variations in response format and latency, and you should work with tenant admins to understand which models are active in your target environments.

How we got here

Microsoft’s relationship with OpenAI has been profoundly expensive and strategically vital. Public reporting puts Microsoft’s committed capital into OpenAI in the low-to-mid double-digit billions. For two years, that investment paid off: Copilot, powered exclusively by GPT models, defined the enterprise AI assistant category.

But as millions of users began generating suggestions, drafts, and transformations daily, the inference cost at scale became a boardroom concern. Running a frontier model for every routine Word autocompletion is like using a supercomputer to do arithmetic. Multiple industry reports indicate Microsoft engineers prioritized cost-efficiency studies in early 2025 that showed routing simpler tasks to smaller models could slash operating expenses without noticeable quality loss.

Simultaneously, Anthropic emerged as a credible enterprise alternative. By mid-2025, Claude Sonnet 4 and Opus 4.1 were available on Amazon Bedrock, offering a safety-focused, production-grade lineup with distinct performance traits. Microsoft’s own A/B tests, according to sources, revealed that Sonnet 4 outperformed in spreadsheet-heavy and visually structured scenarios, while Opus 4 competed well on reasoning tasks.

Commercial hedging also played a role. OpenAI’s growing independence, multi-cloud ambitions, and tough negotiations over terms introduced concentration risk. Adding Anthropic gives Microsoft leverage and a fallback should OpenAI access ever become constrained. And then there’s the legal landscape: Anthropic reached a proposed $1.5 billion settlement with authors over pirated book datasets, a high-profile case that underscored the IP risks of model training. By offering multiple suppliers with clearer contractual safeguards, Microsoft can shield itself and its customers from downstream liability.

What to do now

If your organization uses Microsoft 365 Copilot (or plans to), these are the immediate, practical steps to take:

  1. Update your AI governance policy. Explicitly define which workloads may route to external models, and set the approval workflow for adding new suppliers. Don’t wait until someone turns on Claude by accident.
  2. Review data residency rules. Map where your sensitive data lives and whether sending prompts to an AWS-hosted model breaks any regulatory commitments. If you handle EU personal data, for instance, cross-cloud transfers may require additional safeguards.
  3. Pilot before enabling broadly. Start with a limited group of users and run side-by-side comparisons of OpenAI vs. Claude outputs on your actual corporate documents. Measure not only accuracy but also formatting fidelity, hallucination rates, and refusal patterns.
  4. Instrument your environment. Set up logging that tags every Copilot interaction with the model used, latency, token count, and error status. This is non-negotiable for cost control and auditing.
  5. Negotiate with Microsoft and Anthropic. Push your account teams for clarity on billing mechanics and for contractual terms that specify training-data provenance, data handling guarantees, and indemnification for IP claims.
  6. Communicate with end users. If you enable Claude, inform employees that Copilot’s behavior may shift slightly, and encourage them to flag inconsistencies. Transparency prevents “the AI got worse” complaints.

Outlook

Microsoft’s multi-model move is just the beginning. Over the next 12 to 18 months, expect Copilot Studio to grow into a full marketplace where third-party model providers can offer their own fine-tuned agents. The orchestration layer will become smarter, using real-time A/B signals to route each prompt to the best-performing backend automatically. And Microsoft’s internal MAI models will eventually join the rotation, creating a three-track system: OpenAI for frontier, Anthropic for specialized productivity, and MAI for cost-sensitive scale.

What remains murky: the exact routing rules Microsoft will apply behind the scenes, the final billing structure for third-party inference, and whether the Anthropic author settlement holds up under judicial scrutiny. IT leaders should treat this announcement as an early signal of a multi-year transformation. The Copilot you know today is already evolving into something more like an AI operating system. Governance must evolve at the same speed.