Microsoft's MAI Models Match GPT-5.6 in Excel Tasks at Lower Cost

Microsoft CEO Satya Nadella announced that their in-house MAI series models perform comparably to GPT-5.6 on common Excel tasks while significantly lowering deployment costs. These models, trained in product-specific Reinforcement Learning Environments (RLEs), are integrated into Microsoft's multi-model orchestration system alongside models from OpenAI and Anthropic. This highlights Microsoft's strategic shift toward a hybrid AI orchestration model, reducing reliance on expensive frontier models for specific tasks. By optimizing task-specific models like MAI, enterprises can achieve high performance while dramatically cutting down on infrastructure and operational costs. The MAI models do not require the latest AI accelerators and can run efficiently on older Nvidia H100 and A100 GPUs. They have shown significant improvements in GitHub Copilot and Excel by learning from real customer workflows within specialized reinforcement learning environments.