Microsoft has quietly begun swapping out the AI that powers Copilot inside Excel and Outlook. As of July 7, 2026, a meaningful share of everyday requests in those two apps, such as summarizing email threads, drafting replies, and formatting spreadsheets, is being handled by Microsoft's own MAI models. The move steers high-volume, routine tasks away from OpenAI and Anthropic, and the aim appears to be cutting inference costs. Here is what changed.

A change of engine inside Excel and Outlook

Under the new setup, tens of thousands of Copilot requests in Excel and Outlook are now processed each week by Microsoft's in-house MAI models. These are the bread-and-butter tasks: condensing an email thread, drafting a reply, tidying up the layout of a spreadsheet.

Users do not need to change any settings. Which model handles a request is decided on the server side, so no toggle appears on screen and the experience stays the same as before. MAI is short for Microsoft AI, referring to the family of models Microsoft has built itself.

Not a full break with OpenAI or Anthropic

Microsoft has been clear that it is not cutting ties with its outside partners. More complex work that demands raw capability will continue to run on OpenAI's frontier models. Anthropic's models also remain embedded in specific Office use cases.

In other words, this is closer to a division of labor: high-volume, simple processing goes to MAI, while harder tasks go to the external frontier models. Rather than replacing everything with in-house models, Microsoft has started with the everyday workloads where costs pile up fastest.

What the MAI models are

The MAI family was formally unveiled at Microsoft's developer conference, Build, in June 2026. The lineup includes MAI-Thinking-1 for complex reasoning and MAI-Code-1-Flash for coding, and it spans several areas such as reasoning, coding, image generation, speech, and transcription. Microsoft positions these models as competitive in quality while cheaper to run.

Mustafa Suleyman, who leads Microsoft's AI division, has been candid about the cost logic. He acknowledged that the company pays a great deal to Anthropic, and said the goal is to reduce that burden and ultimately eliminate it.

The goal is inference cost, with ripple effects across the industry

The scale at which Microsoft runs AI is enormous. Because Copilot operates on top of Microsoft 365, used by hundreds of millions of people, the cost of each inference request adds up to a serious sum. Every request routed to an outside provider carries a fee. Preparing in-house models that are good enough for routine work is one of the most direct ways to rein in that spending.

Steering routine requests to cheaper in-house models improves the profit margin on each Copilot seat. For the economics of the business, that is a straightforward positive.

The more pointed question is the effect on OpenAI and Anthropic. Microsoft is OpenAI's largest customer and its primary cloud partner. The more request volume Microsoft shifts to MAI, the more the inference revenue OpenAI earns from that relationship narrows. To dilute this dependency, OpenAI has been expanding direct enterprise deals and consumer products. Anthropic sits within the same dynamic, and the logic that played out in Excel and Outlook, moving work to cheaper in-house models, could well spread to other products.

Summary

The shift sees Microsoft quietly rebalancing the AI stack inside Office, routing everyday work to its cheaper in-house MAI models while keeping OpenAI and Anthropic for the more demanding tasks. It is not a flashy feature launch, but it signals that Microsoft's own models are maturing to a level usable in real work, and that cost discipline is tightening. Whether this division of labor spreads beyond Excel and Outlook is worth watching.