Microsoft is reportedly training sales teams to pitch its in-house MAI models and Microsoft Copilot as faster, cheaper alternatives to offerings from OpenAI and Anthropic—a notable escalation for a company that still sells, hosts, and embeds its partners’ technology. The report, published by Whalesbook, lands days after Bloomberg reported that Microsoft had already moved tens of thousands of weekly AI prompts in Excel and Outlook onto internally developed models.
For Windows and Microsoft 365 customers, the immediate takeaway is not that GPT or Claude are disappearing from Copilot. It is that Microsoft is becoming far more willing to decide which model runs a task behind the familiar Copilot interface—and to make that decision based on cost, latency, reliability, governance requirements, and product integration rather than brand recognition.
That is a material change in posture. Copilot began as a flagship proof point for Microsoft’s OpenAI partnership. In mid-2026, Microsoft is turning it into a model-flexible product layer where OpenAI, Anthropic, and Microsoft’s own MAI models can all be suppliers.
Bloomberg’s July 7 report said Microsoft has started replacing OpenAI and Anthropic models with MAI models in some Excel and Outlook workloads, citing a person familiar with the work. The report did not suggest a complete cutover or claim that one Microsoft model had become universally superior. Instead, it described a targeted shift in high-volume, workplace-oriented tasks where an internally operated model may offer a more economical result.
That distinction matters. An enterprise assistant used for summarizing email threads, extracting action items, rewriting routine messages, or shaping spreadsheet content does not always need the largest or newest frontier model. At Microsoft 365 scale, even a modest per-prompt reduction in inference cost can become financially meaningful when applied across millions of users and repetitive workloads.
The commercial benefit is more than token economics. Microsoft can tune MAI models around its own product behavior, telemetry, security controls, and service architecture. A model that is merely competitive in raw benchmark performance could still be valuable if it produces sufficiently dependable results inside Outlook, Excel, Teams, SharePoint, or Copilot Studio while reducing response time and infrastructure overhead.
Microsoft’s reported sales message follows that logic. Rather than asking customers to buy access to a particular laboratory’s model, the company can sell a governed productivity system: identity through Microsoft Entra, data controls through Microsoft Purview, application context from Microsoft Graph, and user-facing AI through Copilot. The model becomes an interchangeable component—important, but not necessarily the product customers procure.
Microsoft also said it would stop paying a revenue share to OpenAI for products it resells, while OpenAI’s payments to Microsoft continue through 2030, subject to a cap. That arrangement gives both companies more room to pursue their own interests—and removes a structural reason for Microsoft to present OpenAI as the automatic answer to every AI workload.
This is not a clean divorce. Microsoft continues to operate Azure OpenAI Service, continues to license OpenAI technology, and is now preparing to let eligible Microsoft 365 commercial customers use OpenAI-operated models directly through OpenAI as a Microsoft subprocessor. Microsoft added OpenAI to its Online Services subprocessor list on June 23, made the option available on July 9, and plans to enable it automatically for eligible commercial tenants on July 24 unless administrators turn it off.
In other words, Microsoft is simultaneously reducing dependency on OpenAI for certain embedded workloads and expanding customer choice around OpenAI models. That might look contradictory from the outside. It is better understood as Microsoft building leverage: it wants the ability to use a preferred model where it makes sense, without giving up an OpenAI option that many customers specifically want.
Microsoft executive vice president Charles Lamanna told Axios that customers wanted tools that could change the models operating underneath, particularly as leading AI labs continue to leapfrog one another. The company said its Critique approach improved its score on the DRACO deep-research benchmark, although multi-model orchestration comes with a cost and latency penalty.
That is the key constraint on the reported sales push. Microsoft can argue that MAI is efficient and deeply integrated, but it cannot credibly sell an enterprise future in which every task must run on a single in-house system. The company’s own product strategy acknowledges that different models can be useful for different jobs: a smaller Microsoft model for routine productivity tasks, a GPT model for certain advanced Copilot functions, Claude for review or specialized reasoning, and a multi-model workflow for research-heavy work.
The winning pitch, then, is not “Microsoft has replaced OpenAI and Anthropic.” It is “Microsoft can choose the most appropriate model while keeping the experience, data boundary, identity, administration, and billing inside its platform.”
Those figures do not isolate generative AI spending, and they should not be read as a Copilot cost line. But they explain why Microsoft has an incentive to convert infrastructure investment into durable, high-margin software and cloud revenue. Every request Microsoft can serve effectively through an internally optimized model, on capacity it controls, potentially improves the economics of Copilot and Azure.
This is where the sales-force angle becomes important. A new MAI model is not enough by itself to justify the infrastructure buildout. Microsoft needs customers to see Copilot as a business platform that can consolidate AI procurement, reduce shadow use of consumer AI services, and lower the effort needed to deploy AI across Office applications.
The danger is equally clear: if a lower-cost internal route noticeably degrades output quality, users will notice. Knowledge workers can compare results with ChatGPT, Claude, Gemini, or other tools in minutes. Microsoft 365’s distribution advantage makes adoption easier; it does not make weak answers acceptable.
Microsoft says OpenAI-operated models are not currently available in GCC, GCC High, DoD, or sovereign clouds. They are also excluded from in-country processing commitments where applicable, and Microsoft lists several compliance artifacts that are not available for this delivery method, including FedRAMP High authorization, PCI DSS attestation, HITRUST certification, and SOC 1 Type 2 reporting.
That makes provider routing more than a performance setting. It is a data-governance and compliance decision. Organizations should identify which Copilot features rely on OpenAI-operated models, determine whether those features are needed, and document how their tenant policy changes before the automatic enablement date.
Microsoft’s reported competitive campaign is therefore not just a contest with OpenAI and Anthropic. It is a test of whether Copilot can become the control plane for enterprise AI while Microsoft quietly shifts more of the underlying work onto models it owns. The next evidence will be visible in product behavior, admin controls, and whether Excel and Outlook users experience the transition as an improvement—or simply discover that the model beneath Copilot has changed.
For Windows and Microsoft 365 customers, the immediate takeaway is not that GPT or Claude are disappearing from Copilot. It is that Microsoft is becoming far more willing to decide which model runs a task behind the familiar Copilot interface—and to make that decision based on cost, latency, reliability, governance requirements, and product integration rather than brand recognition.
That is a material change in posture. Copilot began as a flagship proof point for Microsoft’s OpenAI partnership. In mid-2026, Microsoft is turning it into a model-flexible product layer where OpenAI, Anthropic, and Microsoft’s own MAI models can all be suppliers.
Excel and Outlook are the first proof of the new operating model
Bloomberg’s July 7 report said Microsoft has started replacing OpenAI and Anthropic models with MAI models in some Excel and Outlook workloads, citing a person familiar with the work. The report did not suggest a complete cutover or claim that one Microsoft model had become universally superior. Instead, it described a targeted shift in high-volume, workplace-oriented tasks where an internally operated model may offer a more economical result.That distinction matters. An enterprise assistant used for summarizing email threads, extracting action items, rewriting routine messages, or shaping spreadsheet content does not always need the largest or newest frontier model. At Microsoft 365 scale, even a modest per-prompt reduction in inference cost can become financially meaningful when applied across millions of users and repetitive workloads.
The commercial benefit is more than token economics. Microsoft can tune MAI models around its own product behavior, telemetry, security controls, and service architecture. A model that is merely competitive in raw benchmark performance could still be valuable if it produces sufficiently dependable results inside Outlook, Excel, Teams, SharePoint, or Copilot Studio while reducing response time and infrastructure overhead.
Microsoft’s reported sales message follows that logic. Rather than asking customers to buy access to a particular laboratory’s model, the company can sell a governed productivity system: identity through Microsoft Entra, data controls through Microsoft Purview, application context from Microsoft Graph, and user-facing AI through Copilot. The model becomes an interchangeable component—important, but not necessarily the product customers procure.
The OpenAI partnership has changed, not ended
The timing is inseparable from Microsoft and OpenAI’s amended agreement announced on April 27, 2026. Microsoft said then that it would retain a license to OpenAI intellectual property through 2032, but that license is now non-exclusive. OpenAI can serve products across cloud providers, while Microsoft remains OpenAI’s primary cloud partner and retains a major shareholder position.Microsoft also said it would stop paying a revenue share to OpenAI for products it resells, while OpenAI’s payments to Microsoft continue through 2030, subject to a cap. That arrangement gives both companies more room to pursue their own interests—and removes a structural reason for Microsoft to present OpenAI as the automatic answer to every AI workload.
This is not a clean divorce. Microsoft continues to operate Azure OpenAI Service, continues to license OpenAI technology, and is now preparing to let eligible Microsoft 365 commercial customers use OpenAI-operated models directly through OpenAI as a Microsoft subprocessor. Microsoft added OpenAI to its Online Services subprocessor list on June 23, made the option available on July 9, and plans to enable it automatically for eligible commercial tenants on July 24 unless administrators turn it off.
In other words, Microsoft is simultaneously reducing dependency on OpenAI for certain embedded workloads and expanding customer choice around OpenAI models. That might look contradictory from the outside. It is better understood as Microsoft building leverage: it wants the ability to use a preferred model where it makes sense, without giving up an OpenAI option that many customers specifically want.
Microsoft has already argued that one model is not enough
Microsoft’s own recent Copilot work points away from a winner-take-all model strategy. In March, Axios reported that Microsoft 365 Copilot Researcher gained a “Critique” layer using Anthropic’s Claude to review answers generated by OpenAI models. Microsoft also introduced a Council option for comparing responses from multiple models.Microsoft executive vice president Charles Lamanna told Axios that customers wanted tools that could change the models operating underneath, particularly as leading AI labs continue to leapfrog one another. The company said its Critique approach improved its score on the DRACO deep-research benchmark, although multi-model orchestration comes with a cost and latency penalty.
That is the key constraint on the reported sales push. Microsoft can argue that MAI is efficient and deeply integrated, but it cannot credibly sell an enterprise future in which every task must run on a single in-house system. The company’s own product strategy acknowledges that different models can be useful for different jobs: a smaller Microsoft model for routine productivity tasks, a GPT model for certain advanced Copilot functions, Claude for review or specialized reasoning, and a multi-model workflow for research-heavy work.
The winning pitch, then, is not “Microsoft has replaced OpenAI and Anthropic.” It is “Microsoft can choose the most appropriate model while keeping the experience, data boundary, identity, administration, and billing inside its platform.”
The data-center bill makes model routing a boardroom issue
The economic pressure behind this strategy is hard to miss. In its fiscal 2025 annual report, Microsoft recorded $64.6 billion in additions to property and equipment, up from $44.5 billion the year before. The company explicitly tied continuing capital expenditure to cloud growth, AI infrastructure, and training.Those figures do not isolate generative AI spending, and they should not be read as a Copilot cost line. But they explain why Microsoft has an incentive to convert infrastructure investment into durable, high-margin software and cloud revenue. Every request Microsoft can serve effectively through an internally optimized model, on capacity it controls, potentially improves the economics of Copilot and Azure.
This is where the sales-force angle becomes important. A new MAI model is not enough by itself to justify the infrastructure buildout. Microsoft needs customers to see Copilot as a business platform that can consolidate AI procurement, reduce shadow use of consumer AI services, and lower the effort needed to deploy AI across Office applications.
The danger is equally clear: if a lower-cost internal route noticeably degrades output quality, users will notice. Knowledge workers can compare results with ChatGPT, Claude, Gemini, or other tools in minutes. Microsoft 365’s distribution advantage makes adoption easier; it does not make weak answers acceptable.
Administrators will need to track the provider boundary
For IT teams, Microsoft’s July 24 OpenAI subprocessor change is an early test of how serious the company is about model choice. Administrators can control access under Copilot settings in the Microsoft 365 admin center and can scope availability by users or Microsoft Entra ID security groups.Microsoft says OpenAI-operated models are not currently available in GCC, GCC High, DoD, or sovereign clouds. They are also excluded from in-country processing commitments where applicable, and Microsoft lists several compliance artifacts that are not available for this delivery method, including FedRAMP High authorization, PCI DSS attestation, HITRUST certification, and SOC 1 Type 2 reporting.
That makes provider routing more than a performance setting. It is a data-governance and compliance decision. Organizations should identify which Copilot features rely on OpenAI-operated models, determine whether those features are needed, and document how their tenant policy changes before the automatic enablement date.
Microsoft’s reported competitive campaign is therefore not just a contest with OpenAI and Anthropic. It is a test of whether Copilot can become the control plane for enterprise AI while Microsoft quietly shifts more of the underlying work onto models it owns. The next evidence will be visible in product behavior, admin controls, and whether Excel and Outlook users experience the transition as an improvement—or simply discover that the model beneath Copilot has changed.
References
- Primary source: Whalesbook
Published: 2026-07-16T00:18:18.327000+00:00
Microsoft Shifts AI Strategy to Challenge OpenAI, Anthropic | Whalesbook
Microsoft is training sales teams to challenge OpenAI and Anthropic, focusing on proprietary AI models to improve efficiency and manage infrastructure costs.www.whalesbook.com
- Official source: blogs.microsoft.com
The next phase of the Microsoft-OpenAI partnership - The Official Microsoft Blog
Amended Agreement Provides Long-Term Clarity The rapid pace of innovation requires us to continue to evolve our partnership to benefit our customers and both companies. Today, we are announcing an amended agreement to simplify our partnership and the way we work together, grounded in...blogs.microsoft.com - Official source: learn.microsoft.com
OpenAI as a subprocessor in Microsoft Online Services | Microsoft Learn
Learn about using OpenAI as a subprocessor in Microsoft Online Services, including how to enable and disable in the Microsoft 365 admin center.learn.microsoft.com - Related coverage: news.bloomberglaw.com
Microsoft Replaces OpenAI, Anthropic With Own AI in Some Apps
Microsoft Corp., looking to reduce AI costs, is starting to replace OpenAI and Anthropic with its own models in software products like Excel and Outlook.news.bloomberglaw.com
- Related coverage: bloomberg.com
- Related coverage: windowscentral.com
"The work we’re doing together remains ambitious": OpenAI ends Microsoft exclusivity in a tone that doesn’t match last week’s leaked memo | Windows Central
OpenAI’s official announcement sounds optimistic, but last week’s leaked memo told a different story.www.windowscentral.com