Microsoft is now answering tens of thousands of Excel and Outlook prompts with its own models

Microsoft has started answering a chunk of its own Copilot prompts with its own models. As of July 7, tens of thousands of prompts a week inside Excel and Outlook are being handled by Microsoft's in-house MAI models instead of the OpenAI and Anthropic systems that used to do that work, according to a Bloomberg report citing people familiar with the change. Microsoft declined to comment.

That is a smaller headline than it sounds, and a bigger one. Smaller because nobody is being cut off: OpenAI's frontier models still handle the harder requests, and Anthropic's models remain wired into specific Office features. Bigger because of what the swapped work actually is.

What moved

The prompts going to MAI are the boring, high-volume ones. Summarize this email thread. Draft a reply. Clean up this spreadsheet's formatting. These are the requests that dominate a productivity app by sheer count, and each one that leaves Microsoft's building carries a per-token bill from whoever answers it.

Route those to a model you own and run yourself, and the bill goes away. What is left going to OpenAI is the smaller set of genuinely hard tasks, where the gap between a frontier model and a cheaper one still shows up in the output.

The MAI models doing the work are the ones Microsoft showed at its Build conference in June, MAI-Thinking-1 and MAI-Code-1-Flash, pitched then as good enough for common tasks and cheaper to operate. "Cheaper to operate" is the whole point here. This is not Microsoft claiming its models beat GPT or Claude. It is Microsoft deciding a lot of Copilot traffic does not need the best model, only an adequate one that costs less per call.

Why now

Microsoft AI chief Mustafa Suleyman said the quiet part in June: "We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost." Earlier reporting had Microsoft on track to spend roughly $500 million a year renting Anthropic models for Office. Once your own inference stack can absorb the routine load, that number stops being a fixed cost and becomes a target.

The math is plain. Microsoft 365 has hundreds of millions of users. At that scale inference is not a rounding error, and every prompt handled in-house instead of rented widens the margin on a Copilot seat. Suleyman has also said a Microsoft model will start doing transcription inside Teams, so Excel and Outlook are unlikely to be where this stops.

What it means for OpenAI

Microsoft is OpenAI's largest customer and its primary cloud partner. When Microsoft serves its own routine prompts, the volume OpenAI bills for through that relationship gets smaller. It is not a break, and OpenAI has spent the past year building direct enterprise and consumer revenue precisely so this kind of shift stings less. But the direction is set: the customer that helped make GPT the default now has an in-house alternative it is actively expanding, and the same logic that justified swapping Excel and Outlook applies to every other product where the task is routine and the volume is high.

None of this required a model breakthrough. It required Microsoft building something merely competent, then doing the arithmetic on what it was paying to rent competence it could own. That is the part worth watching. The frontier gets the press, but the economics of AI are increasingly settled one level down, on the ordinary prompts nobody writes headlines about.