Microsoft unveiled seven new in-house MAI models at its Build event on Tuesday, claiming its flagship systems outperform rivals from Anthropic, OpenAI, and Google. AI CEO Mustafa Suleyman said the reasoning model MAI-Thinking-1 was preferred over Anthropic's Claude Sonnet 4.6 in blind tests by independent evaluators, scored 97% on the AIME 2025 reasoning benchmark, and landed alongside Claude Opus 4.6 on the SWE Bench Pro coding benchmark. The lineup also includes a lightweight coding model for GitHub Copilot, MAI-Image-2.5 (which Microsoft says beats Google's Nano Banana on image editing), a 43-language transcription model, and a speech model that generates voices in 15 languages. Microsoft said its MAI models outperformed GPT-5.5 on quality at 10x lower cost, as it builds proprietary AI systems beyond its longtime reliance on OpenAI.
