'Backstabbing' OpenAI? Microsoft AI Chief: In-House Model Outperforms Peers with Only One-Quarter of the Compute Power!

① Microsoft's head of AI, Suleyman, stated that the company's internally developed new model outperforms similar offerings from OpenAI; ② Microsoft has released the MAI-Base-2 model, which boasts a total of 2.8 trillion parameters, yet its training required only one-quarter of the computational power used by OpenAI; ③ Microsoft is increasingly seeking to reduce its reliance on OpenAI, including by embarking on the development of its own models.

Cailian Press, October 16 (Editor: Huang Junzhi) — Mustafa Suleyman, Microsoft's head of artificial intelligence, told employees that the company's internally developed new model is superior to similar offerings from OpenAI, which is one of the software giant's largest customers and partners.

At an employee meeting held on Thursday, Suleiman stated: "We now have a more powerful pre-trained foundation model whose performance surpasses that of the model currently powering GPT‑5.6 Sol. In fact, the computational resources we used to train this model amount to only one quarter of those employed by OpenAI in training its own model."

Pre-training is the first stage in building artificial intelligence models, during which the model must learn patterns from vast amounts of text, images, and other types of data. This requires substantial, costly computational resources; therefore, if Microsoft can develop a more efficient approach to achieve this goal, it could reduce costs.

Microsoft released this model, named MAI-Base-2, earlier on Thursday. The company stated that it recently completed the model's pre-training. The model employs a mixture-of-experts (MoE) architecture, with a total of 2.8 trillion parameters and 80 billion active parameters.

In a public blog post, Microsoft stated that, across four evaluations, its training method achieved computational efficiency 4.2 to 12.3 times higher than DeepSeek-V4-Pro-Base.

According to Suleiman, Microsoft also conducted private benchmarking tests comparing its model with OpenAI's models, including one codenamed "Spud" that powers GPT-5.6. He added that, in tests measuring AI's ability to understand and predict information patterns, Microsoft's model performed better.

Suleiman also stated that these comparisons are based on the technical information about OpenAI's models that Microsoft possesses.

"Privately, we also compared it with the OpenAI model," Suleiman told employees, while implying that Microsoft had access to proprietary information about the OpenAI model's performance.

Notably, even in private discussions, Suleiman has been deliberately emphasizing the limitations of OpenAI's models. Microsoft has invested tens of billions of dollars in OpenAI and holds a stake in the Venture, valued at potentially over $100 billion. Any criticism of OpenAI could impact the value of this massive investment.

The two companies have also established a significant partnership, with Microsoft offering OpenAI models to its Azure cloud customers. In addition, Microsoft has integrated OpenAI models into its own products, such as Copilot.

However, over the past year or so, the relationship between the two companies has gradually begun to fray. Microsoft has increasingly sought to reduce its reliance on OpenAI, including by embarking on the development of its own models. In March this year, Microsoft established a new joint team for its Copilot products aimed at both enterprises and consumers, and reassigned Suleyman to the company's Superintelligence division, where he now focuses on building AI models.

As of press time, Microsoft declined to comment, and OpenAI did not respond to a request for comment.