Microsoft Launches MAI-Base-2: Claims Performance Surpasses GPT-5.6 Underlying Model with One-Quarter Compute

On October 16, according to a report by Business Insider, Microsoft AI CEO Mustafa Suleyman stated at an employee town hall on the 15th (local time) that a self-developed AI model from Microsoft outperforms OpenAI’s equivalent models, with potentially lower training costs. “Our pre-trained foundation model performs better than the current model powering GPT-5.6 Sol, while actually using only one-quarter of the compute required by OpenAI.”

Pre-training is the first step in AI model development, where models learn patterns from massive datasets of text, images, and other data. Since pre-training requires significant computational resources and is costly, improving training efficiency could allow Microsoft to substantially reduce expenses.

Earlier that day, Microsoft released this self-developed model, named MAI-Base-2.

Suleyman revealed that Microsoft had privately compared MAI-Base-2 against several OpenAI models, including the model codenamed Spud, which underpins GPT-5.6. In tests measuring the ability to understand and predict information patterns, Microsoft’s model achieved superior results.

IT Home learned from the report that the comparison was based on technical documentation regarding OpenAI’s models held by Microsoft. Suleyman noted that Microsoft conducted these comparisons “privately and confidentially,” implying that the company possesses specific insights into OpenAI’s model performance.

Additionally, Suleyman claimed that Microsoft’s new model also holds an advantage in computational efficiency, stating, “For example, we require only one-twelfth of the compute needed by DeepSeek.” However, he did not specify which DeepSeek model was used for the comparison.