Microsoft has begun replacing some OpenAI and Anthropic models in Microsoft 365 applications with its own MAI models for selected tasks in Excel and Outlook, processing tens of thousands of prompts each week. The move signals a strategic shift toward controlling inference costs and operational efficiency as enterprise AI competition pivots from frontier model capability to scalable deployment economics.
Bloomberg reported the deployment, which still accounts for a small portion of Microsoft's overall AI usage. A Microsoft spokesperson declined to comment. The effort is not a break from OpenAI or Anthropic, but a concrete step in a strategy that top executives have been outlining for months.
The shift from frontier models to frontier economics
At Microsoft's Build conference in June, AI Chief Executive Officer Mustafa Suleyman introduced seven new MAI models covering reasoning, coding, transcription, image generation, and more. He said Microsoft wanted to reduce, and ultimately eliminate, spending on Anthropic models. The new MAI-Code-1 model, for instance, delivers coding performance comparable to Anthropic's earlier Opus 4.6 model at a lower operating cost. The new MAI models span reasoning, coding, and transcription, and professionals can learn more through Microsoft AI Courses.
Chief Executive Officer Satya Nadella has publicly argued that long-term AI leadership will depend on building the infrastructure, deployment capabilities, and ecosystems needed to deliver models efficiently. The Bloomberg report suggests Microsoft is now executing that strategy inside its own products.
A portfolio of models approach
The changes reflect an architectural shift becoming common across enterprise AI platforms. Rather than relying on a single foundation model, vendors assemble portfolios of models optimized for different tasks. Complex reasoning still requires the most capable frontier models from OpenAI or Anthropic. Routine activities-email assistance, spreadsheet analysis, summarization, document generation-can often be handled by smaller, less expensive models without a noticeable difference for end users.
For every Copilot interaction, Microsoft consumes computing resources including inference tokens, GPU capacity, networking, memory, storage, and safety systems. As enterprise adoption grows, even small reductions in per-request cost translate into substantial operational savings. The use of MAI models for tens of thousands of weekly prompts, while still a fraction of total usage, shows the company is beginning to redirect workloads to its own models where it makes economic sense.
Why this matters for IT and development
This shift signals that cost-efficient deployment is becoming as critical as model performance in enterprise AI. IT teams will need to design and manage multi-model environments where prompt routing matches workload requirements to the most cost-effective model. For developers integrating AI into Office workflows, the change means less reliance on expensive third-party APIs for routine tasks, and a growing need to understand how in-house models perform on specific jobs. The economics of inference are moving from a fixed cost per request to a managed portfolio of models, and that will reshape procurement, architecture, and optimization decisions.
Your membership also unlocks:
