Microsoft's in-house AI models begin replacing OpenAI in production

Microsoft is moving aggressively to build an AI strategy that doesn't hinge on OpenAI, rolling out internally developed models into its core productivity tools and seeing strong early results.

Building in-house AI muscle#

The software giant has quietly begun deploying proprietary AI models within its own ecosystem. An internal model now running in Excel achieves performance comparable to GPT-5.6 for routine user tasks while demanding significantly less infrastructure spend, according to Microsoft's own claims. Meanwhile, the company's MAI-Code-1-Flash model—developed and refined within GitHub Copilot—is already running at scale across "millions of developers" handling daily coding work.

The push reflects a pragmatic business calculation: Microsoft wants to deliver AI capabilities to its massive customer base without funneling all the resulting revenue to OpenAI and Anthropic, its two key portfolio companies and external partners.

Performance and efficiency gains#

Microsoft's metrics suggest its models are holding their own. The MAI-Code-1-Flash variant shows "approximately 10% higher code accept rate" compared to GPT 4 Mini and Claude Haiku 4.5 when tested in Visual Studio Code, with notably better user retention rates. The efficiency angle matters equally: these models achieve results while consuming fewer tokens than similarly sized competitors.

On the Excel front, Microsoft appears less focused on frontier performance and more interested in delivering practical, cost-effective solutions to ordinary users who don't scrutinize AI leaderboards.

The Microsoft advantage#

Building and refining these models in-house gives Microsoft a structural edge rivals can't easily replicate. Unlike pure AI research labs, Microsoft sits atop a staggering wealth of enterprise coding patterns, workflow data, and real-world usage signals flowing through Excel, Teams, Copilot, and other services. The company can feed this proprietary information directly into its "hill-climbing machine"—its term for the iterative improvement cycle—without any risk of proprietary data leaking to outside labs.

This data flywheel, once it spins efficiently, could allow Microsoft to improve rapidly and narrow the capability gap with frontier models far faster than conventional training cycles allow.