Microsoft Claims New AI Model Outperforms OpenAI Using 75% Less Computing Power

Microsoft has announced a significant development in its artificial intelligence research, with AI chief Mustafa Suleyman claiming that the company’s new model, MAI-Base-2, outperformed a comparable OpenAI model while requiring substantially less computing power to train. Speaking during an internal town hall meeting on Thursday, October 15, 2026, Suleyman told employees that Microsoft had developed a stronger pretrained model using approximately one-quarter of the computing resources OpenAI used for its comparable system. The announcement signals Microsoft’s growing ambition to compete directly with leading AI developers rather than depending primarily on technology supplied by external partners. However, the performance claims were based on internal comparisons, and Microsoft has not yet publicly released sufficient independent evaluation data to establish that the model outperforms OpenAI’s systems across a broad range of tasks.

The new model, MAI-Base-2, was developed by Microsoft’s AI division as part of its broader effort to build advanced models in-house. According to Suleyman, the system performed better than a comparable OpenAI model associated with GPT-5.6, which he referred to by the internal codename “Spud.” He also said Microsoft’s model demonstrated greater training efficiency than DeepSeek-V4-Pro-Base, a model developed by the Chinese AI company DeepSeek. These comparisons are important because training advanced AI models requires enormous amounts of computing power, electricity, specialised hardware and technical expertise. If Microsoft can consistently achieve comparable or better results with fewer resources, it could reduce development costs and make it easier to train increasingly capable systems. Nevertheless, the claims should be treated as Microsoft’s reported assessment until detailed benchmarks and independent testing become available.

One of the most significant aspects of the announcement is the potential reduction in computing costs. Training a large AI model involves processing enormous quantities of data through specialised processors over extended periods. The process can require thousands of advanced chips operating in large data centres, creating substantial infrastructure and energy expenses. Using only one-quarter of the computing resources required by a competing model could provide Microsoft with a considerable economic advantage if the comparison holds under equivalent training conditions. However, computing efficiency does not depend solely on the number of processors used. Factors such as model architecture, training data, optimisation techniques and the amount of computation allocated to evaluation can influence the results. Microsoft will therefore need to provide additional technical information before the broader significance of MAI-Base-2’s reported efficiency can be fully assessed.

The announcement also reflects changing dynamics in Microsoft’s relationship with OpenAI. Microsoft has invested heavily in OpenAI and incorporated its models into products and services across its technology ecosystem. At the same time, the company has been expanding its own AI research capabilities to gain greater control over model development, performance, cost and deployment. Developing competitive in-house models could give Microsoft more flexibility when building Copilot features, enterprise software and autonomous AI agents. It could also reduce its dependence on a single external provider as competition intensifies among OpenAI, Anthropic, Google, DeepSeek and other AI developers. Microsoft has already been investing in custom AI chips, including its Maia accelerator family, as part of a wider strategy to improve the economics of running advanced AI systems.

Despite the promising claims, MAI-Base-2’s long-term impact will depend on how it performs when made available for broader testing and real-world use. A model that demonstrates strong pretraining results may still require additional development to perform reliably in coding, mathematical reasoning, scientific research, conversational assistance and complex business workflows. Developers and enterprise customers will also evaluate its safety, reliability, operating costs and compatibility with existing applications. Microsoft has not yet established through publicly available independent evaluations that MAI-Base-2 is universally superior to competing models. Even so, the announcement highlights an important trend in the AI industry: companies are increasingly competing not only to build more capable models, but also to train and operate them more efficiently. If Microsoft can substantiate its claims, MAI-Base-2 could strengthen its position in the global AI market and intensify competition among the industry’s leading technology companies.