Microsoft AI CEO Says MAI-Base-2 Beats OpenAI GPT With Less Compute

Microsoft released its self-developed model MAI-Base-2 on October 15. On the same day, Microsoft AI CEO Mustafa Suleyman stated at an all-hands meeting that internal tests showed the model outperforms the base model currently used for GPT-5.6 Sol in benchmarks evaluating the understanding and prediction of information patterns; he also noted that Microsoft used only one-quarter of the pre-training compute compared to OpenAI’s corresponding model. These comparisons with OpenAI stem from Suleyman’s internal remarks, and Microsoft’s publicly available technical articles do not provide complete comparative data for external verification. According to Microsoft’s disclosures, MAI-Base-2 is a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active parameters. The company stated that, in its evaluations, the model exhibits higher pre-training computational efficiency than DeepSeek-V4-Pro-Base. However, MAI-Base-2 is currently still undergoing post-training, so the performance of the base model cannot be directly equated with the final user-facing product experience. While maintaining investment and product partnerships with OpenAI, Microsoft is increasing its investment in self-developed models. The release of MAI-Base-2 indicates that Microsoft is attempting to gain control over more stages of the pipeline from model training to product application, though its cost and performance advantages remain to be validated by subsequent public testing.