Microsoft is transitioning several of its Copilot products, including those integrated into Excel and Outlook, by replacing AI models developed by OpenAI and Anthropic with proprietary in-house alternatives. This strategic move aims to significantly reduce the company’s overall expenditures on external artificial intelligence services.
Adoption of Internal MAI Models
According to reports from Bloomberg, Microsoft’s internal MAI models are already actively processing tens of thousands of requests each week within Excel and Outlook. Previously, these specific applications were heavily reliant on the capabilities provided by OpenAI and Anthropic. While the homegrown models currently handle only a modest portion of total queries, Microsoft plans to continue reducing spending on third-party AI services over time. Furthermore, the MAI technology is slated for integration into GitHub Copilot, and a specialized transcription model is expected to launch in Teams soon.
Performance Claims Versus Benchmarks
During the Build conference, Microsoft introduced seven new artificial intelligence models, including MAI-Thinking 1, which was presented as its first dedicated reasoning model. While Microsoft asserted that Thinking-1 could rival Sonnet 4.6 and Opus 4.6 in coding ability based on human evaluations, benchmarks published at the time revealed a different picture. The data indicated that Thinking-1 significantly lagged behind its competitors from OpenAI and Anthropic, performing only roughly on par with Deepseek V3.2.
Impact on Consumers and Pricing Models
This shift suggests that consumers using Copilot and Office services might receive AI capabilities of potentially lower performance for the same subscription cost, thereby aiding Microsoft’s cost reduction goals. Mustafa Suleyman, Microsoft’s head of AI, publicly confirmed this strategy in June: “We pay a lot of money to Anthropic—so our goal is to reduce and ultimately eliminate that cost.” This statement contrasts with earlier statements from the company where it advocated for platform neutrality and criticized vendor lock-in when dealing with OpenAI and Anthropic.
Additionally, Microsoft CEO Satya Nadella suggested that billing practices for AI services could evolve away from fixed, flat-rate subscriptions toward a usage-based pricing structure. One potential implementation involves making the cheaper MAI models the default choice, while requiring customers to pay a surcharge if they wish to access third-party options from OpenAI or Anthropic.
Data Training and Licensing Concerns
Microsoft has stated that its proprietary MAI models are trained using clean data derived solely from commercial licenses, ensuring safety for enterprise users. However, technical documentation reveals that Microsoft utilized the Common Crawl dataset. This dataset consists of freely accessible web content, and its legal status regarding use in AI training remains undetermined. Although this practice is common among other major AI corporations, Microsoft has emphasized that its own data sources are particularly clean.
