When Satya Nadella told the Build 2026 audience in San Francisco that "the time has come for every company to move from consuming a frontier model to fully participating at the frontier," he was talking about his customers. He was also, quietly, talking about Microsoft itself. The company has spent $13 billion investing in OpenAI and another $5 billion in Anthropic. At Build 2026, it unveiled seven new AI models it built entirely on its own, and called the family MAI.
The seven models span the full range of tasks Microsoft's products need: MAI-Thinking-1 handles reasoning and complex multi-step problems, MAI-Code-1-Flash is built into GitHub Copilot, MAI-Image-2.5 generates and edits images, MAI-Transcribe-1.5 converts speech to text across 43 languages, and MAI-Voice-2 produces natural-sounding synthetic speech. All were trained from scratch by Microsoft's own teams, with no outputs borrowed from OpenAI's models. This last point was made explicitly, repeatedly.
The simplest explanation for all of this is cost. Microsoft AI CEO Mustafa Suleyman said the company had tailored MAI-Thinking-1 for consulting firm McKinsey and achieved ten times better cost efficiency than using OpenAI's GPT-5.5 for the same tasks. Running your own models on Azure, rather than paying a third party per token, is an obvious financial lever, especially as AI usage scales inside Microsoft 365, Windows, and Azure itself.
But cost is only part of what is happening. MAI-Thinking-1 is Microsoft's first serious bid to compete in the reasoning model category that OpenAI, Google, and Anthropic have been racing to lead. It has 35 billion active parameters inside a mixture-of-experts architecture with a 256,000-token context window. Microsoft says it outperforms Claude Sonnet 4.6 in blind human evaluations. PCMag's hands-on reviewer was less impressed, finding it limited by the absence of real-time internet access, and concluded that its performance did not obviously justify choosing it over existing alternatives. That gap between benchmark claims and real-world utility is a recurring feature of frontier model announcements, not specific to Microsoft.
What is specific to Microsoft is the position this puts the company in relative to OpenAI. The two are still deeply entangled: OpenAI models remain available on Azure, GitHub Copilot has long run on OpenAI's Codex, and neither company has said the partnership is unwinding. But MAI-Code-1-Flash is being built into the same GitHub Copilot that Codex powers. MAI-Image-2.5 sits alongside DALL-E in Microsoft's design tools. The relationship has shifted from dependency to something more like strategic competition inside an ongoing alliance.
For enterprise buyers, the picture is more pragmatic than dramatic. The ability to fine-tune a model on proprietary data without routing it through a third-party provider is worth something on its own. Suleyman's McKinsey example is the clearest illustration of the value proposition: companies that can tailor the model to their domain can get substantially better cost-to-performance ratios than those using general-purpose frontier APIs at list price.
The MAI models are in limited or early preview. Several require access requests, and MAI-Thinking-1 is listed as "coming soon" for most users. Microsoft's build conferences have a history of bold demos that arrive gradually, and the gap between keynote and widespread availability is real. Still, the direction is now clear. The company that backed two of its most important AI partners with billions of dollars has decided that the next phase of AI development requires building something of its own.