Microsoft releases MAI-Base-2, claiming it beats OpenAI at a quarter of the compute

On October 15, 2026, Microsoft released MAI-Base-2, a pretrained foundation model built on a mixture-of-experts (MoE) architecture with 2.8T total parameters and 80B active parameters. The company published a blog post earlier that day. Microsoft says its training method is 4.2x to 12.3x more compute-efficient than DeepSeek-V4-Pro-Base; Techmeme relayed Microsoft's claim as 4x to 12x and noted that the figures have not been independently verified. Microsoft also referenced MAI-Thinking-1, a general-purpose foundation model post-trained on top of MAI-Base-1 that it announced in June.

According to Business Insider reports cited by East Money and ITHome, Mustafa Suleyman, CEO of Microsoft AI, told employees at a Thursday all-hands meeting that MAI-Base-2 outperforms the model currently powering GPT 5.6 Sol, while using only one quarter of the compute OpenAI used to train its model. Suleyman said Microsoft compared MAI-Base-2 privately against several OpenAI models, including one codenamed Spud that underpins GPT-5.6, and that Microsoft's model scored better on tests of AI comprehension and prediction of information patterns. ITHome reported that the comparison drew on OpenAI model technical materials in Microsoft's possession, and Suleyman described the comparison as conducted "privately and confidentially."

The situation currently rests at Microsoft's own announcement and internal statements: the model has been released and pretraining completed, but the efficiency multiples and performance comparisons remain Microsoft's own claims. The reports do not mention third-party verification or any response from OpenAI.