Microsoft AI Model Beats OpenAI With 1/4 the Compute
According to Business Insider, Microsoft AI CEO Mustafa Suleyman told employees on October 15 that the company's newly developed AI model outperforms a comparable model from OpenAI while requiring significantly less computing power.
At an employee meeting on Thursday, Suleyman said Microsoft's pretraining model performs better than the model currently powering GPT-5.6 Sol, while using just one-quarter of the computing resources OpenAI used to train its model. Pretraining is the initial stage of AI development, during which models learn patterns from large amounts of text, images and other data. Improving training efficiency could significantly reduce AI development costs.
Microsoft announced its new model, MAI-Base-2, earlier on Thursday, confirming that pretraining had recently been completed. The model uses a mixture-of-experts (MoE) architecture, with 2.8 trillion total parameters and 80 billion active parameters.
Suleyman said Microsoft had privately benchmarked MAI-Base-2 against OpenAI models, including a model codenamed "Spud" that powers GPT-5.6. According to him, Microsoft's model performed better on tests measuring AI's ability to understand and predict patterns in information. He said the comparisons were based on technical information Microsoft had obtained about OpenAI's models.
The claims come as Microsoft's relationship with OpenAI evolves. Microsoft has invested billions of dollars in OpenAI and integrates its models into Azure cloud services and products such as Copilot. However, Microsoft has increasingly pursued its own AI development to reduce its reliance on OpenAI.
Suleyman also claimed that MAI-Base-2 requires only one-twelfth of the computing resources used by a comparable DeepSeek model. Microsoft said in a public blog post that the model's pretraining efficiency exceeds that of DeepSeek-V4-Pro-Base.
In March, Microsoft reorganized its Copilot teams serving enterprise and consumer customers and moved Suleyman to its superintelligence division, where he focuses on developing AI models.
