Microsoft Claims Its New MAI-Base-2 Beats the Model Powering GPT-5.6 While Using a Quarter of the Training Compute

Microsoft AI CEO Mustafa Suleyman told employees on Thursday that the company's newest in-house base model outperforms the model powering OpenAI's GPT-5.6 while using a quarter of the training compute, according to Business Insider. The comments came hours after Microsoft announced the model, called MAI-Base-2, earlier the same day.

"We now have a stronger pre-trained base model than the model that currently powers GPT-5.6 Sol," Suleyman said during a town hall meeting. "And in fact, we used one-quarter of the compute that OpenAI used to train their model."

Microsoft says MAI-Base-2 is a mixture-of-experts model with 2.8 trillion total parameters and 80 billion active ones, and that pretraining was recently completed. Suleyman said Microsoft privately benchmarked the model against OpenAI's models, including one nicknamed "Spud" that powers GPT-5.6, and that it scored better on tests measuring how well AI systems understand and predict patterns in information.

He added that Microsoft's model is also more compute-efficient than rivals' open models, saying it needs 12 times less computation than DeepSeek's. A public Microsoft blog post described the pretraining comparison as being against DeepSeek-V4-Pro-Base. All of the performance and efficiency figures above are Microsoft's own claims and have not been independently verified.

The claims are pointed given the relationship between the two companies. Microsoft has invested billions in OpenAI and holds a stake likely worth more than $100 billion, and it still resells OpenAI models through Azure and uses them in Copilot. But the partnership has frayed over the past year as Microsoft pushes to reduce its dependence on OpenAI, including building its own frontier models under Suleyman's superintelligence unit.

Suleyman suggested Microsoft has special access to information about how OpenAI's models perform, saying the comparisons were made "privately and confidentially" using technical information Microsoft holds about OpenAI's systems. Microsoft declined to comment and OpenAI did not respond to a request for comment, Business Insider reported.

The efficiency framing matters as much as the benchmark claims. If a 2.8-trillion-parameter model can match a frontier rival at a fraction of the training cost, the economics of proprietary pretraining start to look different — for Microsoft's own margins, and for an industry currently absorbing a rough repricing of AI-linked debt. Whether independent benchmarks bear out the town hall rhetoric is the question Microsoft has yet to answer.