MAI-Thinking-1: Microsoft enters the advanced-reasoning AI race with its own from-scratch model

The Microsoft AI (MAI) team has just announced MAI-Thinking-1, its first advanced reasoning model developed in its own laboratories. Microsoft claims that the AI model possesses reasoning capabilities superior to those of Claude’s Sonnet 4.6, citing blind human side-by-side evaluations. However, this statement comes directly from Microsoft and not from an independent third-party tester, so some scepticism is warranted. MAI-Thinking-1 is part of a family of seven AI models, marking a milestone in Microsoft’s AI strategy.

The new models not only cover reasoning, but also image, voice, transcription and code, forming an interconnected ecosystem designed to work on real-world tasks. These are:

• MAI-Thinking-1, the flagship reasoning model. This is trained from scratch with clean data, does not depend on or use third-party models, and competes with the best AI models for software engineering and mathematical reasoning.
• MAI-Code-1-Flash, intended for efficient programming. Designed especially for GitHub Copilot, VS Code, and Microsoft, it has 5 billion active parameters and is more economical than equivalent models.
• MAI-Transcribe-1.5, which Microsoft claims is the best transcription model in the world, boasting that it is five times faster than comparable models and available for transcriptions in 43 languages.
• MAI-Image-2.5, with quality image generation, reported to be superior in Arena scoring compared to Nano Banana Pro.
• MAI-Voice-2, natural audio and voice generation in 15 languages and which will be available at a more accessible price.

One of the most notable aspects of the announcement is the “Frontier Tuning”, through which AI models are adapted to the specific needs of each company, using their own information and workflows in a secure environment. This way, business data is incorporated into the model and remains under the customer’s control. As a preliminary result, Microsoft reports that the MAI model tuned for Excel matches GPT 5.4 and is up to 10 times more efficient.

MAI concludes its announcement by indicating that the ultimate objective is to achieve “Humanist Superintelligence”, namely advanced AI systems designed to serve people and organizations, not to replace them.

More information can be found in Microsoft’s announcement here.