Microsoft unveiled seven new AI models at its Build developer conference on Tuesday, highlighted by the 35-billion-parameter MAI-Thinking-1. Described by Microsoft AI lead Mustafa Suleyman as its "first reasoning model," MAI-Thinking-1 reportedly surpasses Anthropic's Claude Sonnet 4.6 in overall quality and achieved a 97% on the AIME benchmark and 53% on SWE Bench Pro for coding tasks. A key selling point is that MAI-Thinking-1 was trained from the ground up without distillation, addressing enterprise concerns about legal uncertainties and copyright infringement related to training data sourcing. This approach, emphasizing clean and commercially licensed data lineage, aims to build trust for production deployment. The announcement also includes image-generation models (MAI-Image-2.5 and Flash), transcription (MAI-Transcribe-1.5), speech generation (MAI-Voice-2 and Flash), and a coding model (MAI-Code-1-Flash). This marks Microsoft's largest AI product launch since its independently developed MAI-Voice models in August, signaling a strategic shift away from its heavy reliance on OpenAI.