Satya Nadella outlines Microsoft's MAI strategy as Excel model matches GPT-5.6 at lower cost

Microsoft CEO Satya Nadella has outlined the company's evolving AI strategy, arguing that advances in software have brought the industry to a point where AI capabilities can be delivered at significantly lower cost while still matching the performance of frontier models for many real-world tasks.

In a post on X, Nadella questioned how the benefits of frontier AI models could be diffused across the broader ecosystem. He said the key lies in optimizing the cost-to-outcome frontier in real-world contexts by using the right model for the right task. According to Nadella, this philosophy underpins Microsoft's in-house MAI family of models.

Microsoft said the MAI models have been built from the ground up with a clean data lineage and are optimized to transfer learning from general-purpose models to specialized enterprise reinforcement learning environments. Rather than relying exclusively on large frontier models, the company aims to deliver comparable capabilities at lower cost through models tailored for high-volume enterprise workloads, while reserving frontier models for more demanding use cases.

More from Tech

As part of this strategy, Microsoft highlighted two new deployments of its MAI models inside its own products, with specialized agentic workloads now powering GitHub Copilot and Microsoft Excel.

The company said MAI-Code-1 Flash, which was introduced in GitHub Copilot in June, is already being used by millions of developers for day-to-day coding tasks. According to Microsoft, the model has outperformed other similarly sized models while using fewer tokens, making it both faster and more efficient.

Testing across different domains

Microsoft said Excel served as an important test of MAI-Code-1 Flash's ability to perform outside the coding domain it was originally trained for. The company used the model to transition from agentic coding tasks to agentic knowledge work, training it through reinforcement learning on spreadsheet tools and Excel-based workflows.

These efforts resulted in a model with a strong understanding of Excel workflows that is both more efficient and less expensive to run. According to Microsoft, production feedback indicates that the MAI model delivers performance on par with GPT-5.6 for the most common Excel tasks while requiring significantly lower deployment costs, as it can run on both Nvidia H100 and A100 GPUs instead of relying solely on the latest AI accelerators.

Microsoft said the results reinforce its broader strategy of developing smaller, task-specific AI models by leveraging its full product stack, including models, agents, runtime infrastructure, and product-specific evaluations. The company is now extending this approach beyond GitHub Copilot and Excel to other agentic products, including Copilot Chat, Outlook, and PowerPoint.