微軟推出自主研發的MAI模型，以取代OpenAI和Anthropic在Excel和Outlook中的應用

Microsoft has begun integrating its proprietary MAI series models into core Office applications like Excel and Outlook, phasing out reliance on external providers such as OpenAI and Anthropic. Internal sources indicate that tens of thousands of AI prompts are now handled by these models weekly. While still a small fraction of total usage, this shift signals significant progress in Microsoft’s ability to build competitive, cost-effective AI infrastructure.

Sharma noted, "We are saving a substantial amount annually, with the goal of reducing—and eventually eliminating—our reliance on external providers."

This strategy is driven by clear economic logic. Microsoft consumes vast quantities of AI tokens through tools like Copilot, currently benefiting from a discounted long-term partnership with OpenAI. However, as these preferential terms diminish, Mustafa Sharma’s team is accelerating in-house development to avoid future high costs from leading AI labs.

During this year’s Build Developer Conference in June, Microsoft unveiled seven new AI models. One reportedly matches the encoding capabilities of Anthropic’s Opus 4.6 but at a lower cost. Sharma openly stated, "We’ve spent heavily on Anthropic, so our objective is to cut these costs and ultimately remove them entirely." The MAI model has already demonstrated its cost-efficiency in specific use cases, from Excel data processing to Outlook email assistance.

From Office to Teams, in-house models are now fully integrated

The adoption of MAI extends beyond the Office suite. Sharma announced that Microsoft’s self-developed speech recognition model will be integrated into Teams’ video conferencing and other products within the next few months. Additionally, the MAI model powers GitHub Copilot, supporting frequent office tasks such as code assistance and meeting note-taking.