Microsoft Takes Over Excel with In-house Model: Common Task Performance Close to GPT-5.6, Lower Cost

According to Insight Beating monitoring, Microsoft AI has announced progress on two products with specialized language models. MAI-Code-1-Flash has been used in GitHub Copilot, while another unnamed model has been running in Excel.

Microsoft stated that the Excel model performs close to GPT-5.6 on common tasks. It has a smaller footprint, can be deployed on H100 and A100, and has lower operational costs.

MAI-Code-1-Flash has been utilized by millions of developers. Its code acceptance rate is about 10% higher than GPT-5.4 Mini and Claude Haiku 4.5, with a 10% lower Token consumption.

Microsoft is also training Copilot Chat, Outlook, and PowerPoint specialized models using the same approach.