Microsoft has reportedly begun replacing the OpenAI and Anthropic models behind some features in apps like Excel and Outlook with its own in-house "MAI" models. This is not a wholesale migration yet — the company is only slotting its own models into situations where cost or data handling makes them the better fit — but it stands out as a gradual move to loosen a dependence Microsoft spent years building.

Its own models are now running inside Excel and Outlook

According to reports, tens of thousands of AI prompts each week in Excel's spreadsheet features and Outlook's email-related functions are already being handled by Microsoft's in-house MAI models. Even so, this is not the kind of change that happens all at once. Across Microsoft's Copilot assistant as a whole, the bulk of actual processing is still handled by OpenAI and Anthropic models, with MAI stepping in only where conditions such as cost efficiency or data residency line up.

Running the models on its own Azure infrastructure lets Microsoft hold down the fees it pays to outside providers. The company says one MAI model, tuned for the consulting firm McKinsey, outdid OpenAI's GPT-5.5 on cost efficiency by a factor of ten. Rather than chasing flashy benchmark wins, the thinking is that with reach into hundreds of millions of Office and Teams users, shifting even a slice of that traffic to its own models moves serious money.

Seven MAI models unveiled at Build 2026

MAI is being nurtured as a family rather than a one-off model. At its Build 2026 developer conference, Microsoft unveiled seven MAI models spanning reasoning, coding, image generation, voice, and transcription. The transcription-focused "MAI-Transcribe-1" is being tested across Teams and Copilot, while the "MAI-Image-2" image generator is being worked into Bing and PowerPoint. In GitHub Copilot, developers can now choose Microsoft's own models for AI-assisted coding. The company plans to extend its in-house speech transcription model to Teams and other products going forward.

A rethink of the exclusive contract created a "three-way hedge"

Behind Microsoft's push into in-house models is a shift in its relationship with OpenAI. Under the earlier contract, Microsoft was restricted from independently pursuing frontier AI. But when the two companies renegotiated in 2025, the exclusive arrangement was dissolved, freeing Microsoft to build competing models of its own. It kept its license to use OpenAI's technology through 2032, with only the leash removed. The rework cut both ways: OpenAI can now sell its own models through rivals such as AWS.

As a result, Microsoft holds three options at once — a large stake in OpenAI, Anthropic's "Claude" embedded in Copilot, and its own MAI models — a "three-way hedge." Rather than betting everything on a single choice, it has built a structure that lets it match the best model to each use case.

Why the rush to build in-house now

CEO Satya Nadella has reportedly worried that leaning too heavily on a single partner could turn Microsoft into "the next IBM." The idea is that owning the model, rather than renting it, is the way to avoid that fate. Mustafa Suleyman, who leads the AI division, also said in June that the company was trying to cut its spending on Anthropic by making more use of MAI.

The bill for outside AI is far from small, and every feature brought in-house trims that invoice. The OpenAI partnership that once symbolized Microsoft's AI strategy is being quietly reworked, one swapped-in model at a time.

Summary

Microsoft has started switching some Excel and Outlook features from outside models to its own MAI, focusing the changeover on cases where cost efficiency and data management favor it. This is not a full migration, and OpenAI and Anthropic still power the core of Copilot, but by leveraging its reach to hundreds of millions of users, Microsoft is clearly aiming to shift weight toward its own models bit by bit. As a hedge against relying on any single company, how far MAI spreads across Office and Teams will be worth watching.