TechFlow reports on March 7, according to The Information, that a team led by Kalkidan Simeneh, deputy to Microsoft AI head Mustafa Suleyman, has recently completed a series of model trainings internally referred to as MAI. Sources indicated these models perform nearly on par with OpenAI and Anthropic's leading models in widely recognized benchmark tests.
The report notes that the Microsoft team is also training reasoning models using chain-of-thought techniques, which allow AI systems to "think through" problems before solving them—enabling Microsoft to compete directly with OpenAI. Microsoft has reportedly begun testing the replacement of OpenAI's models in Copilot with the much larger MAI models, surpassing earlier Phi models in scale.
In addition, insiders revealed that Microsoft is evaluating alternative models for Copilot, including offerings from Anthropic, Musk's xAI, DeepSeek, and open-source models from Meta Platforms. This initiative signals Microsoft’s efforts to reduce reliance on OpenAI and position its in-house models to directly compete with similar API-based products from OpenAI and other AI labs.
