Microsoft Expands Model Choice with Anthropic in Copilot and GitHub

Microsoft’s product teams have quietly flipped a strategic switch: after years of tightly integrating OpenAI models across Azure, Bing and Windows, the company is now offering Anthropic models as first-class options inside Microsoft 365 Copilot and other developer tools — a clear signal that model diversification has moved from pilot to production inside Microsoft’s AI strategy.

Background​

Microsoft’s multi‑billion dollar relationship with OpenAI rewrote the economics and product roadmap for enterprise AI. Over the past several years Microsoft has committed tens of billions of dollars of funding and exclusive cloud infrastructure arrangements to host OpenAI’s models, while shipping those capabilities into Microsoft 365, GitHub, Bing and Windows experiences. That partnership remains foundational to Microsoft’s AI portfolio — but the company’s recent moves show it is deliberately broadening the choices that customers and developers can use inside the Microsoft ecosystem.
Two developments crystallized this shift in September 2025. First, Microsoft announced that Anthropic’s Claude Sonnet 4 and Claude Opus 4.1 are now available as selectable models inside Microsoft 365 Copilot (in Researcher and Copilot Studio), giving enterprise customers a direct alternative to OpenAI models for deep reasoning and agent orchestration. Second, GitHub Copilot Chat had already started letting developers select Anthropic and Google models alongside OpenAI’s offerings — a move Microsoft first introduced in late 2024 and has since expanded. These changes formalize multi‑model choice as a Microsoft product principle.

---

What changed and why it matters​

Microsoft’s integration of Anthropic models in Copilot is not a mere UI tweak — it reflects several important strategic realities:

• Model choice as product-level capability. Customers can now pick models by capability (reasoning, creativity, safety), cost, or regulatory profile within the same Copilot workflow, rather than stitching together outputs from multiple vendors manually.
• Vendor diversification for resilience. Relying on a single model supplier introduces supply‑chain, pricing and political risk; adding Anthropic, Google and other partners reduces that exposure.
• Competitive sourcing of best-in-class features. Different models excel at different tasks; Microsoft is optimizing user outcomes by letting organizations choose a model that best fits the job (for example, research vs. code generation vs. summarization).
• Cloud hosting complexity. Anthropic’s models are currently hosted outside Microsoft‑managed environments (notably on AWS), which raises operational questions around latency, data residency and contractual terms for enterprise customers. Microsoft’s blog notes Anthropic models are hosted outside Microsoft-managed environments and subject to Anthropic’s terms.

These are not incremental product choices; they are structural changes that will reshape how IT leaders govern, secure and measure generative AI inside corporate estates.

---

What Microsoft announced (technical specifics)​

Which Anthropic models, where they appear​

Microsoft’s official Microsoft 365 blog post states that Claude Sonnet 4 and Claude Opus 4.1 are available in:

• The Researcher agent inside Microsoft 365 Copilot (a multi‑step reasoning assistant that reads across email, files, chats and web data).
• Copilot Studio, where developers and integrators design, mix and orchestrate agents for specific enterprise workflows.

Researcher users can toggle between OpenAI and Anthropic models; Copilot Studio exposes these models in a dropdown so agents can call different models for specialized tasks. Microsoft also says availability begins via opt‑in programs for customers who want to evaluate the new model options.

Hosting and data handling​

Microsoft disclosed that Anthropic models used in Copilot are hosted outside Microsoft‑managed environments and operate under Anthropic’s own terms and conditions. This has immediate implications for enterprise compliance and security programs: data flows, retention, access controls and contractual liability must be reviewed before enabling external models for sensitive workloads.

---

Cross‑checking the market context​

A handful of concurrent industry moves underscore why Microsoft is diversifying:

• GitHub Copilot multi‑model support — Microsoft added Anthropic and Google models as alternatives in GitHub Copilot Chat in late 2024, a precedent showing that Microsoft’s developer tooling would not be limited to OpenAI alone.
• OpenAI’s infrastructure deals — OpenAI has expanded partnerships and capital commitments with several vendors, notably a major multi‑gigawatt infrastructure agreement with NVIDIA that includes up to $100 billion of phased investment tied to hardware deployment. That deal is intended to secure GPU supply for OpenAI’s scaling plans and is among the largest private‑sector infrastructure commitments in AI to date.
• Broadcom and Oracle signals — Reports indicate Broadcom has secured multi‑billion dollar orders for custom AI chips (analysts pointing to OpenAI as a potential customer), while Oracle and others are also scaling infrastructure to meet the industry’s massive compute needs. Some headlines have placed Oracle‑OpenAI contracts or commitments in the hundreds of billions when aggregated into broader projects like “Stargate,” but those numbers often combine multi‑party spending estimates and should be treated as macro projections rather than fixed bilateral contract values.

Together, these developments show that the AI value chain — from models to chips to datacenters — is being restructured rapidly, and hyperscalers are adapting their product strategies accordingly.

---

Strategic rationale: why Microsoft is doing this​

1) Reduce single‑supplier risk without burning bridges​

Microsoft’s commercial relationship with OpenAI has been deep and reciprocal, ranging from sizable funding to preferential model access and joint engineering. But the AI market is evolving faster than any single partnership can cover. By adding Anthropic models, Microsoft hedges against supplier disruptions, pricing shocks, or changing commercial terms while preserving essential OpenAI integrations. This is prudent supply‑chain hygiene at planetary scale.

2) Compete on product flexibility, not exclusivity​

Enterprise customers value control and choice. Microsoft can differentiate its Copilot offerings by being the platform where organizations can unify, orchestrate and manage multiple model providers from a single admin plane, instead of forcing teams to run siloed POCs across multiple vendors. That positioning strengthens Microsoft’s platform lock‑in in a different — arguably more sustainable — way.

3) Capture workloads that favor different model architectures​

Anthropic emphasizes safety and instruction‑following alignment in certain scenarios; OpenAI often leads in scale and ecosystem integration. Giving customers the ability to route tasks to the best model for the job (e.g., Claude for high‑assurance reasoning vs. GPT variants for creative synthesis) can measurably improve outcomes for enterprise workflows.

---

Risks, tradeoffs and open questions​

The new multi‑model approach introduces important tradeoffs that enterprise IT, legal and security teams must confront.

Data governance and compliance​

Because Anthropic’s models are currently hosted outside Microsoft‑managed environments, enabling them creates trans‑service data flows that many compliance programs find difficult to justify without contract amendments and rigorous data‑processing addenda. Data residency, encryption in transit and at rest, and legal process for law enforcement requests are all practical concerns that IT teams must evaluate before opting in. Microsoft has flagged this explicitly in its communication.

Latency and performance consistency​

Routing some requests to third‑party hosts (AWS, Anthropic) and others to Azure or Microsoft‑hosted models can produce variable latency, throttling characteristics, and cost profiles. For tightly coupled workflows — large document analysis, real‑time assistants, or low‑latency code completions — this heterogeneity will matter. Enterprises should benchmark end‑to‑end performance for critical processes.

Licensing, revenue share and the economics of exclusivity​

Microsoft’s historic investment and revenue‑sharing relationship with OpenAI — which industry reporting pegs at around $13 billion in commitments to date — created financial incentives and exclusivities that shaped product integrations. Diversifying to other model vendors threatens none of that legal reality, but it does change the economics: using Anthropic models may be priced differently and carry different contractual obligations. Reported investment totals and revenue‑share terms vary across coverage, and some public figures remain estimates rather than finalized, audited disclosures. Treat dollar figures and percentage splits as indicative, not definitive, unless confirmed in regulatory filings or company disclosures.

Fragmentation and developer complexity​

Giving developers and product teams more options improves raw capability but increases operational complexity. Observability, cost attribution, model evaluation, and retraining of prompt engineering practices will all require new governance — otherwise organizations risk model sprawl and inconsistent outputs across teams.

---

Technical implications for Windows, Microsoft 365 and Azure​

Windows and Copilot on device​

Microsoft’s recent work to integrate generative AI into Windows and Microsoft 365 has emphasized tight coupling between local experiences and cloud models. Adding Anthropic models into Copilot means Windows‑based workflows that call Copilot (desktop apps, Outlook, Teams) may now involve external host calls — increasing the importance of local policy enforcement (e.g., preventing sensitive PII from being transmitted) and of admin controls that can restrict which models are available for a tenant or OU.

Azure’s positioning as a neutral model catalog​

Microsoft’s messaging suggests Azure is becoming an orchestration layer for models — including those hosted elsewhere — through Copilot Studio and the Azure Model Catalog. That repositioning helps Azure retain control as enterprises centralize governance while still allowing third‑party model innovation to flourish. Over time, Microsoft may also seek hosting agreements to run Anthropic models on Azure if customers demand in‑Microsoft hosting to simplify compliance.

---

The broader industry picture: chips, datacenters and megadeals​

Microsoft’s diversification occurs against a backdrop of enormous infrastructure commitments across the AI stack. Notable developments that shape Microsoft’s options and OpenAI’s choices include:

• NVIDIA’s strategic partnership with OpenAI to deploy at least 10 gigawatts of NVIDIA systems, backed by up to $100 billion in phased investment tied to deployments — a deal that locks in GPU supply and capital to fuel model training at scale. This reshapes how raw compute is procured and which companies can scale models rapidly.
• Broadcom’s announced $10 billion customer order for custom AI chips, widely reported and speculated to involve OpenAI as a customer. If confirmed, this signals OpenAI’s multi‑supplier chip strategy and potential move toward custom silicon.
• Oracle, SoftBank and other infrastructure plays augment the compute and datacenter landscape through projects that aggregate vast capital commitments; public reporting on total dollar figures can conflate multi‑party, multi‑year infrastructure spending with single‑party contracts, so care is needed when quoting headline sums.

Taken together, the compute layer is diversifying just as much as the model layer — and those two shifts are mutually reinforcing.

---

What enterprises should do now (practical checklist)​

Enterprises that deploy Microsoft 365 Copilot, GitHub Copilot, or are considering Copilot Studio should move deliberately:

• Inventory Copilot touchpoints. Map where Copilot and model-driven features are used across apps, automations and integrations.
• Set admin guardrails. Use Microsoft 365 admin controls to restrict which models are allowed per organizational unit; require opt‑in and review for high‑risk teams.
• Contractual review. Update data processing agreements and vendor risk assessments for scenarios where models are hosted outside Microsoft‑managed infrastructure.
• Benchmark performance and cost. Run A/B tests between OpenAI and Anthropic models for your most valuable workflows and measure accuracy, latency and total cost of ownership.
• Improve observability. Instrument calls to AI models with centralized logging, cost tags and output verification to prevent hallucination‑driven errors from propagating into business processes.

These steps convert the promise of model choice into an operational capability rather than a compliance headache.

---

Strengths of Microsoft’s multi‑model approach​

• Flexibility and customer choice — Organizations can align model selection to policy, cost and outcome requirements without leaving Microsoft tools.
• Faster innovation cadence — Microsoft can surface models that outperform for special use cases more quickly, giving customers immediate benefit.
• Platform differentiation — Microsoft shifts from being a single‑vendor integrator to a model orchestration provider, strengthening platform value even as it reduces supplier lock‑in.

---

Weaknesses and potential risks​

• Compliance friction — External hosting creates friction for regulated industries and will require painstaking contract work.
• Operational complexity — Multi‑model deployments necessitate new governance, increasing the burden on IT and SRE teams.
• Perception risk with OpenAI relationship — While Microsoft continues to partner closely with OpenAI, public moves toward competitors can create perception issues with partners and investors that favor clarity over ambiguity. Market reporting about investment totals and revenue shares has already drawn investor attention.

---

What remains uncertain (flagged claims)​

Certain headline numbers and future projections require cautious treatment:

• Reported totals such as Microsoft’s cumulative investment in OpenAI (commonly cited around $13 billion) and revenue‑share mechanics are reported by multiple outlets but vary by accounting treatment and public disclosure. Treat these as aggregated figures gleaned from multiple reporting threads rather than precise audited numbers.
• Large infrastructure project price tags (for example, multi‑hundred billion dollar figures connected to “Stargate” or Oracle/OpenAI contracts) are often aggregate estimates across partners and years and should not be read as single bilateral contract values unless confirmed. Several reputable outlets report large aggregate spending plans, but the headline sums mix private investments, projected infrastructure costs, and multi‑party commitments. Flag these as macro estimates rather than fixed contract amounts.
• NVIDIA’s up‑to‑$100 billion phased investment is part of a letter of intent and is contingent on deployment milestones; it is accurate to describe the commitment but important to note it is staged and conditional.

---

Bottom line​

Microsoft’s decision to expose Anthropic models inside Microsoft 365 Copilot and developer tooling is a defining moment in enterprise AI: it converts model selection from a vendor negotiation into a product feature. For businesses, that choice is a powerful lever to optimize outcomes and manage supplier risk — but it comes with measurable governance, contractual and operational costs.
Enterprises should treat this as an opportunity to formalize AI governance, benchmark model outputs, and build the telemetry and contractual guardrails needed to make multi‑model deployments safe and predictable. Microsoft, meanwhile, is repositioning Azure and Copilot as the orchestration layer that can unify a splintering model ecosystem. That bet — turning platform ubiquity into platform neutrality for models — could preserve Microsoft’s central role in enterprise software even as the AI supplier landscape fragments and scales.

---

Quick reference: immediate checklist for IT leaders​

• Inventory Copilot usage and sensitive data touchpoints.
• Turn on admin controls before enabling external models.
• Require DPA updates and vendor risk assessments for Anthropic workloads.
• Pilot Anthropic models in low‑risk workflows and measure outcomes against OpenAI models.
• Centralize logging and model output verification workflows.

Adopting these steps will turn the strategic advantage of model choice into operationally safe gains for businesses that rely on Microsoft’s productivity stack.
Conclusion: Microsoft’s move to incorporate Anthropic marks a mature phase in enterprise AI — one where choice, governance and orchestration matter at least as much as raw model capability. The next 12–24 months will show whether customers prefer a single‑vendor depth of integration or Microsoft’s vision of a neutral, multi‑model Copilot that lets organizations pick the right brain for each problem.

---

Source: PYMNTS.com Microsoft Turns to Anthropic in Shift From OpenAI Relationship | PYMNTS.com