Microsoft’s Copilot Cost Problem Comes Home to Excel and Outlook
Microsoft’s recent decision to route tens of thousands of AI prompts from Excel and Outlook to its proprietary MAI models instead of OpenAI’s GPT or Anthropic’s Claude each week marks a strategic shift in how the company powers Copilot, signaling a move toward in-house AI infrastructure and away from expensive third-party model dependence. This is no experimental pilot: according to reporting based on people familiar with the work, these MAI models have moved beyond testing and are now handling real production workloads inside Microsoft 365. While external models still process most Copilot traffic today, the direction of travel is clear. Microsoft is actively unwinding a dependency it spent years and billions building, and the reason is frank and unapologetic—Copilot cost reduction, not model loyalty.
Mustafa Suleyman’s Mandate: Reduce and Ultimately Eliminate External AI Spend
The driver behind Microsoft’s MAI push is not subtle: “The driver is simple: cost.” Mustafa Suleyman, CEO of AI models at Microsoft, has said the company is routing more workloads away from Anthropic to MAI specifically to cut spending. At the June Build conference, he went further, stating that Microsoft’s plan is to “reduce and ultimately eliminate” the amount of money it sends to Anthropic. This fits a broader pattern. Earlier this year, Microsoft reduced workers’ access to external tools like Claude Code after costs skyrocketed, then moved GitHub Copilot to token-based billing and removed employee access to Claude Fable over data-retention concerns. Across the industry, token costs have spiraled to the point where one company reportedly burned through half a billion dollars in a single month, and another used its entire annual AI budget in three months. Microsoft’s response is clear: own more of the stack or pay for it forever.
MAI Models: From Backup Plan to OpenAI Alternative Inside Copilot
Microsoft is not trying to beat every benchmark; it is trying to build a viable OpenAI alternative within its own ecosystem. At Build, the company unveiled seven MAI models spanning reasoning, coding, image generation, speech, and transcription. The lineup includes MAI-Thinking-1 for complex reasoning, MAI-Code-1-Flash for software development, MAI-Image-2.5, MAI-Voice-2, and MAI-Transcribe-1.5. One MAI model tuned for consulting firm McKinsey reportedly beat OpenAI’s GPT-5.5 on cost efficiency by a factor of ten, underscoring the economic logic behind this shift. Another coding model is said to match Anthropic’s Opus 4.6 programming capabilities at lower cost. MAI models already appear in GitHub Copilot and Copilot for Business and Enterprise users, where they sit alongside GPT and Claude. The goal is a multi-model Copilot and Azure AI platform where tasks can be routed dynamically to MAI, GPT, or Claude depending on what makes the most financial and technical sense.
In-House AI Infrastructure: Strategic Control Over an Expensive Future
Microsoft’s reach across hundreds of millions of Office and Teams seats means that even small routing changes—from OpenAI and Anthropic to first-party MAI models—translate into meaningful savings. Pushing more Copilot workloads onto Microsoft AI models makes costs easier to manage, even if the return on investment is still hard to quantify. The company is turning Copilot and Azure AI into multi-model platforms that choose the best model per task, but the economic bias is clear: privilege in-house AI infrastructure whenever acceptable quality and reliability can be met. A renegotiated OpenAI deal now lets Microsoft build competing models while keeping a license to OpenAI’s technology through 2032, while OpenAI can also sell through rivals. That mutual freedom changes the risk calculus. Satya Nadella reportedly feared Microsoft becoming “the next IBM” if it leaned too hard on a single AI partner; MAI is the hedge against that future.
What Comes Next for Office AI: A Copilot That Serves Microsoft First
The next wave of Microsoft AI models will not be abstract research projects; they will be baked directly into frontline productivity tools. Suleyman has said a Microsoft-built transcription model will roll out to Teams and other products in the coming months, and the company has already announced plans for a Teams-based transcription feature. MAI models are already servicing tens of thousands of prompts in Excel and Outlook, a small but growing share of overall Microsoft 365 AI workloads. Historically, those workloads favored ChatGPT and Claude, but the funnel is now tilting toward MAI. Microsoft is not ending its partnerships with OpenAI or Anthropic; customers will still be able to tap MAI, GPT, or Claude within the same ecosystem. But make no mistake: every new Excel AI feature, every Teams transcript, every coding suggestion in Copilot will be evaluated through a single lens—does it keep Microsoft paying external model providers, or does it help eliminate that cost over time?
