Microsoft has started replacing the AI that powers Copilot features in Excel and Outlook with its own internally-built models, called MAI. The scale of the shift has not been announced publicly: it came out via Bloomberg, which reported that the MAI models are already processing tens of thousands of requests per week in those apps. Previously, both applications relied more heavily on AI from OpenAI and Anthropic.
The motivation is straightforward. Microsoft's head of AI, Mustafa Suleyman, said in June: "We pay a lot of money to Anthropic, so our goal is to reduce and ultimately eliminate that cost." He also called Anthropic "extremely expensive" and said many companies are looking for cheaper options. This is not subtle: a senior executive of one of the world's largest companies publicly listed cost reduction on a named supplier as a corporate goal.
The background context matters here. Microsoft has a discounted deal with OpenAI that runs through 2032, so OpenAI's costs are manageable for now. Anthropic is a different story: Microsoft has no such sweetheart deal there, making it the more expensive dependency and the obvious first target for replacement.
The MAI models now cover a wide range of tasks. Microsoft launched seven of them at its developer conference in June: a reasoning model, a coding model, an image generator, a voice model, and a transcription tool. The transcription model is expected to be added to Teams soon. The coding model is already available inside GitHub Copilot.
The performance question is where this gets uncomfortable for Copilot customers. Microsoft claimed its new reasoning model matches Anthropic's Opus 4.6 on coding tasks. But independent benchmarks put it behind both Anthropic and OpenAI's current generation by a wide margin, roughly on par with a Chinese open-source model called DeepSeek V3.2. The comparison Microsoft made was against an older Anthropic model that Anthropic itself has since surpassed with newer releases. Suleyman acknowledged the gap, saying Microsoft has "closed an enormous gap in six months" while admitting that Anthropic has since moved further ahead.
For anyone paying for Microsoft 365 Copilot or a similar subscription, the practical question is: will the AI get worse without the price going down? The honest answer right now is: possibly, yes, for some tasks. The MAI models still handle only a small share of overall requests, so the impact is limited today. But Microsoft's stated direction is to expand their use over time.
There is also a pricing change looming. Microsoft's CEO Satya Nadella has hinted at a shift from flat monthly subscriptions toward pay-per-use billing. One likely outcome: the cheaper MAI models become the default, and users who want the better models from OpenAI or Anthropic pay a surcharge on top. Microsoft would effectively pass its own AI costs onto customers rather than absorbing them.
One more detail worth knowing: Microsoft has marketed its MAI models as being trained on "clean, commercially licensed data," a claim designed to appeal to regulated industries worried about legal exposure. The technical documentation, however, shows Microsoft used the same pool of freely scraped web content that every other AI company uses, and whose legal status for AI training is not settled. Every AI company does this. Microsoft just markets itself as if it does not.
For businesses currently using Microsoft 365 Copilot, the near-term action is simple: pay attention to whether the quality of AI-assisted tasks in Excel, Outlook, and Teams changes over the coming months. If you notice it getting less reliable on complex tasks, this is why. And if Microsoft introduces tiered pricing that charges extra for better AI, that will be the moment to reassess whether the flat-rate subscription still makes sense for your team.