Microsoft launches MAI-Thinking-1 reasoning model at Build, beats Anthropic's Sonnet

At its annual Build conference, Microsoft AI unveiled seven new models, led by MAI-Thinking-1, its first reasoning model. The 35-billion-parameter model was trained on enterprise-grade, commercially licensed data to address copyright concerns. Microsoft claims it beat Anthropic's Claude Sonnet 4.61 in blind reviewer evaluations and matched Claude Opus 4.6 on the SWE Bench Pro coding benchmark. MAI-Thinking-1 is designed for multi-step agentic tasks and is available in private preview on Microsoft Foundry.

Alongside the reasoning model, Microsoft introduced MAI-Code-1, an ultra-efficient coding model tuned for GitHub, coming to Copilot and VS Code. MAI-Image-2.5 (and its flash variant) surpassed Nano Banana Pro on ELO ratings and hit #3 on the LM Arena Leaderboard; it's live in PowerPoint, Foundry, and rolling out in OneDrive. New voice and transcription models (MAI-Voice-2 with 15 more languages than its predecessor, and MAI-Transcribe-1.5 with 43-language coverage) round out the release. Suleyman emphasized that all models are watermarked by default and offer cost improvements up to 10x versus competitors. Additionally, Microsoft announced a collaboration with Mayo Clinic to develop a frontier healthcare model.