Microsoft MAI Models Replace OpenAI and Anthropic Inside Excel and Outlook

by Zaira Varghese 10. July 2026 Microsoft 0

For the first time, Microsoft MAI models — not OpenAI’s, not Anthropic’s — are quietly completing tens of thousands of AI prompts every week inside Excel and Outlook. The swap happened without a keynote, without a press release, and without most users noticing a thing.

Bloomberg first reported the shift, and the strategic logic is not subtle. Microsoft AI chief Mustafa Suleyman has been unusually blunt about the motive: “We pay a lot of money to Anthropic — so our goal is to reduce and ultimately eliminate that cost.”

What the Microsoft MAI swap actually changes

The models are not going after the hard problems. Microsoft is targeting what you might call the commodity layer of AI work — the tasks that are enormously high in volume and modest in difficulty:

Drafting routine email replies in Outlook
Summarizing long message threads
Generating simple spreadsheet formulas in Excel
Cleaning up and reformatting text

These are the prompts that make up the overwhelming bulk of Copilot traffic. They are also the prompts where a frontier model is wildly overqualified — the AI equivalent of hiring a structural engineer to hang a picture frame. Every one of them, until now, has been billed to Microsoft at frontier-model rates.

The economics behind the decision

Microsoft has committed tens of billions to OpenAI and, more recently, signed a substantial capacity agreement with Anthropic to serve Copilot workloads. Those contracts made sense when nobody else could deliver the quality. They make considerably less sense when your own research division can serve an email summary at a fraction of the marginal cost.

Inference is now a line item, not a research budget. At Copilot’s scale, shaving even a fraction of a cent off each prompt compounds into hundreds of millions annually. Microsoft has said the in-house models will expand to GitHub Copilot and Teams next — the two products with the highest prompt volume left on the board.

What it means for OpenAI and Anthropic

Neither lab is losing Microsoft as a customer. Both remain the destination for genuinely hard reasoning, long-context analysis, and agentic work. But the revenue mix is changing in a way that should worry anyone modeling AI lab economics on volume.

The commodity tier was always the ballast — high volume, predictable, and unglamorous. If the biggest software distributor on earth pulls that ballast in-house, what remains for the labs is the premium tier: lower volume, higher margin, and far more exposed to the next model release. It is a structurally less comfortable position.

There is also a competitive wrinkle. Microsoft distributes OpenAI’s technology and simultaneously builds a substitute for it. Suleyman’s team, staffed heavily with ex-DeepMind and ex-Inflection researchers, is not building a toy.

Should users care?

Practically, most won’t notice. Microsoft says the routing is quality-gated — if a prompt looks complex, it goes to a frontier model. The Microsoft MAI systems handle the easy majority. In theory, users get identical output at lower cost to Microsoft.

In practice, silent model substitution deserves scrutiny. Users who chose Copilot partly because it ran on GPT-class models were not told when that stopped being reliably true. The routing logic is opaque, and “quality-gated” is a claim, not a measurement anyone outside Redmond can verify.

The bigger pattern

This is what commoditization looks like from the inside. The first phase of the AI boom rewarded whoever had the best model. The second phase rewards whoever owns distribution and can vertically integrate the parts of the stack that stop being differentiated.

Microsoft owns the applications where hundreds of millions of people already work. Once a capability becomes routine enough to build yourself, owning the surface beats renting the intelligence. The Microsoft MAI rollout is the first real proof that the calculation has flipped — and it almost certainly will not be the last time a hyperscaler runs the numbers and reaches the same conclusion.