Microsoft AI CEO Mustafa Suleyman told staff at an October 15 town hall that MAI-Base-2, the in-house model Microsoft announced that day, outperforms the pre-trained model behind OpenAI’s GPT-5.6 Sol while using a quarter of the training compute, according to Ashley Stewart at Business Insider. Suleyman said Microsoft ran that test privately using technical details it holds on OpenAI’s models, but Microsoft’s public post only compares MAI-Base-2 with open-weight rivals such as DeepSeek-V4-Pro-Base. The model is still in post-training, and Microsoft plans to license its weights to select partners through Foundry. Microsoft owns an OpenAI stake probably worth over $100 billion and resells OpenAI’s models on Azure, even as it builds its own to depend on them less.
