According to Bloomberg, Microsoft has started using internally developed MAI series models in its major Office products such as Excel and Outlook, replacing the models from OpenAI and Anthropic that it previously relied on. According to sources, tens of thousands of AI prompts are now processed by the MAI models each week. Although the overall proportion is still small, this change indicates that Microsoft has made substantial progress in developing competitive and low-cost AI models.

Sharma stated, "We are saving a huge amount of money every year, and we aim to reduce our dependence on external sources and eventually eliminate it completely."

This is clearly based on a clear economic calculation. Microsoft consumes a massive amount of AI tokens through work-assisting tools like Copilot, and currently benefits from a long-term partnership with OpenAI to obtain the latest technology at discounted prices. However, this preferential arrangement is shrinking, and Mustafa Sharma's team aims to accelerate in-house development to avoid being charged exorbitant fees by top-tier AI research labs in the future.

At the Build Developer Conference in June this year, Microsoft announced seven new AI models all at once. One of them is said to have similar encoding capabilities to Anthropic's previous popular model Opus 4.6 but can be implemented at a lower cost. Sharma candidly stated at the time, "We have paid a lot of money to Anthropic, so our goal is to reduce this cost and ultimately eliminate it completely." From data processing in Excel to email assistance in Outlook, the MAI model has proven its cost-effectiveness in specific scenarios one by one.

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

The scope of MAI model adoption is not limited to the Office trio. Sharma revealed that Microsoft's self-developed speech recognition model is planned to be introduced into Teams' video conferencing application and other products within the next few months. At the same time, the MAI model is also available in GitHub Copilot, covering multiple frequent office scenarios such as code assistance and meeting note-taking.