Microsoft's MAI image models are moving into the products where OpenAI used to be the obvious default, and that tells you more about AI economics than any partnership statement can.
The swap is no longer a lab story or a conference demo. It is product plumbing now. Microsoft's own MAI image models are already powering image generation in Bing and PowerPoint, according to Microsoft's fiscal third-quarter remarks, and the company has since pushed MAI-Image-2.5 into PowerPoint while rolling it out to OneDrive. That is the story.
Microsoft introduced MAI-Image-2 in March and said it had reached the top three model families on Arena.ai's text-to-image leaderboard. By June, the company was promoting MAI-Image-2.5 and MAI-Image-2.5-Flash as part of a seven-model MAI lineup announced at Build 2026, alongside MAI-Thinking-1, MAI-Code-1-Flash, MAI-Transcribe-1.5, MAI-Voice-2, and MAI-Voice-2-Flash. Arena's public leaderboard now shows Microsoft AI's MAI-Image-2.5 among the leading image models, while Microsoft's own Build notes said the model reached No. 2 on Arena's single-image editing benchmark as of June 2. That is enough proof of intent.
Bloomberg reported on July 7 that Microsoft had begun replacing OpenAI and Anthropic models with its own MAI models in Excel and Outlook to reduce AI costs. Tens of thousands of weekly prompts in those apps are now being completed by Microsoft-built models, Bloomberg reported, citing a person familiar with the work. Microsoft's own April earnings call had already given you the same direction in plainer corporate language: the company said it was focused on first-party model work to differentiate its copilots and reduce cost of goods sold. The pattern is not subtle.
The cost argument is the real argument
Every token Microsoft routes to its own models is a token it does not pay a partner to process. When Microsoft sells a Microsoft 365 Copilot subscription and the underlying inference runs on a third-party model, some of the economics leave the building. Swap in a MAI model running on Azure, and more of that cost stays inside the house. The bill moves inside.
Microsoft has been unusually direct about this. In its Build 2026 notes, the company projected a tenfold improvement in output tokens per dollar from a fine-tuned MAI model compared with GPT-5.5, based on public GPT pricing and Microsoft's own MAI pricing data. You do not need to treat an internal projection as gospel to understand the business logic. Direction matters more.
The partnership has changed too. In an April 27 joint announcement, Microsoft and OpenAI restructured the deal on several fronts: OpenAI can now serve its products across any cloud provider, the IP licence runs through 2032 on a non-exclusive basis, and Microsoft will no longer pay a revenue share. OpenAI then announced on July 9 that GPT-5.6 would become the preferred model for Microsoft 365 Copilot across Word, Excel, PowerPoint, Chat and Cowork. Preferred is not exclusive. TechCrunch made the useful point at the time: that label does not undo Bloomberg's reporting that Microsoft is relying more on its own models to cut costs.
The lesson for startups is uncomfortable
OpenAI still matters to Copilot. Microsoft continues to sell OpenAI models through Azure, and GPT-5.6 is now positioned inside Microsoft 365 Copilot rather than outside it. But if you are building an AI startup on top of one model provider, you should pay attention. Your most important customer may eventually become your most obvious competitor.
Microsoft is the easiest case to see because its OpenAI relationship is so visible and its investment was so large. Once internal quality gets close enough, the margin argument becomes hard to resist. Speed helps too. Microsoft says MAI-Image-2-Efficient is 22% faster and four times more efficient than MAI-Image-2 under its April test conditions, and priced 41% lower for image output tokens than the flagship model. Those are not philosophical claims. They are procurement claims.
For AI startups that sell access to models, the riskier version of this story is not simply Microsoft replacing OpenAI in parts of Copilot. Do not make this only about Microsoft. A company processing enough AI queries will eventually ask whether API dependency still makes sense at its scale. Some will build. Most will not have to, because the hyperscalers - Microsoft, Google, Amazon and the rest - are already pushing the model layer toward lower prices and easier substitution.
OpenAI is not going away from Copilot tomorrow. The product bets Microsoft is making now, in PowerPoint slides, Bing images, Outlook drafts and Excel prompts, are accumulating into something real. The $13 billion bought access and a head start. It did not buy permanence.
Also read: Nvidia's Jetson chips are becoming the default AI layer for Moon missions • Black Forest Labs collapses image, video, and robot control into a single model with FLUX 3 • Etched is raising at $10 billion and $20 billion simultaneously and investors are lined up
