Microsoft Claims MAI-Base-2 Beats OpenAI's GPT-5.6 Model With a Quarter of Training Compute

Microsoft AI chief Mustafa Suleyman told employees during an all-hands meeting on Oct. 15 that the company's in-house pretrained foundation model, MAI-Base-2, outperforms the OpenAI model currently powering GPT-5.6 Sol, while using only a quarter of the training compute.

"Our pretrained foundation model is performing better than the model currently powering GPT-5.6 Sol, and we actually used a quarter of the compute OpenAI used for training," Suleyman said, according to Business Insider.

Microsoft released MAI-Base-2 earlier the same day. Suleyman said the company had privately compared MAI-Base-2 with several OpenAI models, including the model codenamed Spud that supports GPT-5.6, and that Microsoft's model scored better in tests measuring AI's ability to understand and predict patterns in information. The comparisons were based on technical documentation of OpenAI models that Microsoft possesses, he said, suggesting the company has special insight into OpenAI's model performance.

Suleyman also highlighted the new model's compute efficiency, citing DeepSeek as an example: "We require only one-twelfth of the compute." He did not specify which DeepSeek model the comparison referred to.