Microsoft launches its own AI model, says it beats OpenAI with a quarter of the training power

Pre-training is the first step in building an AI model. The model learns patterns from huge amounts of text and image data, which takes lots of computing power and costs a lot. Better training efficiency means much lower costs.

Microsoft AI CEO Mustafa Suleyman told employees that the company's own model, MAI-Base-2, beats OpenAI's similar products in performance. It used only a quarter of OpenAI's computing power for training. The model was released the same day.

Suleyman said Microsoft privately compared MAI-Base-2 with several OpenAI models, including one codenamed Spud that powers GPT-5.6. Microsoft's model scored better in tests on understanding and predicting patterns.

He added that the new model is more efficient too, needing only one-twelfth of the computing power compared to DeepSeek, though he did not say which DeepSeek model was used.