Microsoft AI head: New basic model surpasses OpenAI similar products with only one-quarter of its computing power

Abstract:

According to "Business Insider", Microsoft AI business CEO Mustafa Suleyman told employees,

A model developed independently by Microsoft is better than a similar model provided by OpenAI. Currently, OpenAI is one of Microsoft's largest customers and partners.

Suleiman

Suleiman said at a staff meeting on Thursday: "We now have a more powerful pre-trained base model that outperforms the model currently driving GPT 5.6 Sol. In fact, we trained this model with only a quarter of the computing power that OpenAI used to train its model."

Pre-training refers to the first stage of building an AI model. In this stage, the model learns patterns from massive amounts of text, images, and other information. This process requires a lot of expensive computing power, so if Microsoft finds a more efficient training method, it can save costs.

Microsoft announced the model, called MAI-Base-2, earlier Thursday

Microsoft announced it as MAI-Base-2

Suleiman said that Microsoft privately tested this model against OpenAI's models, including a model code-named "Spud" that drives GPT-5.6. He added that Microsoft's model performed better in tests that measure AI's ability to understand and predict patterns in information.

Suleiman said these comparisons are based on technical information about the OpenAI model that Microsoft has. "We also compared it to OpenAI's model privately and confidentially," Suleiman told employees, hinting that Microsoft was able to obtain special information about the performance of the OpenAI model.

It is worth noting that even in private discussions, Suleiman emphasized the limitations of the OpenAI model. Microsoft has invested billions of dollars in OpenAI and holds a stake in the startup that could be worth more than $100 billion. Any criticism of OpenAI could affect the value of this massive investment.

The two companies have also entered into an important partnership, with Microsoft providing OpenAI models to Azure cloud service customers. Microsoft also uses OpenAI models in its own products such as Copilot.

But over the past year or so, cracks have gradually emerged in the relationship between the two companies

Suleiman also said that from the perspective of computing power usage efficiency, Microsoft's new AI model requires only one-twelfth of the calculations required by DeepSeek. Suleiman did not specify which DeepSeek model, but the company stated in a public blog post that the pre-training efficiency of MAI-Base-2 is higher than that of a similar DeepSeek product called DeepSeek-V4-Pro-Base.

In March of this year, Microsoft formed a new joint team for its Copilot products for enterprises and consumers, and transferred Suleiman to the company's superintelligence department to focus on building AI models.

As of press time, Microsoft has declined to comment, and OpenAI has not responded to requests for comment.