Microsoft builds its own AI model to reduce OpenAI reliance

Microsoft unveiled MAI-Thinking-1, its first self-developed artificial intelligence reasoning model, at the company's annual Build developer conference in San Francisco. The model represents a significant milestone in Microsoft's strategy to reduce its dependence on OpenAI, whose public market debut is imminent, and demonstrates the tech giant's ability to build competitive AI systems from the ground up. MAI-Thinking-1 is a medium-sized model with 1 billion parameters, 35 billion of which are active, and features a 256 KB context window designed for high efficiency and low token costs. According to Microsoft, the model performs on par with advanced solutions like Claude Opus 4.6 in certain benchmarks and outperformed Claude Sonnet 4.6 in blind user tests. The AI was trained from scratch using clean, licensed data within a fully self-sufficient end-to-end chain, without relying on distillation from existing OpenAI models or external hardware and software resources. The model particularly excels in mathematical reasoning and software development tasks. MAI-Thinking-1 is part of a broader family of seven new AI models announced by Microsoft, including MAI-Code-1-Flash for agentic programming tasks, MAI-Image-2.5 for text-to-image and image-to-image generation, MAI-Voice-2 for voice synthesis, and MAI-Transcribe-1.5 for transcription supporting 43 languages. The MAI-Thinking-1 model is currently available as a private preview on Microsoft Foundry, with a public release planned for the MAI Playground platform. CEO Satya Nadella emphasized that the release marks a significant shift, stating that companies should now move from merely consuming frontier models to actively participating in the frontier ecosystem.