Microsoft used its Build 2026 developer conference on June 2 to launch seven homegrown AI models under its MAI brand, a clear push toward model self-sufficiency after years of leaning on partner OpenAI. The lineup is led by MAI-Thinking-1, the company's first flagship reasoning model, a 35-billion active-parameter system that Microsoft says delivers high efficiency at low token cost and matches Anthropic's Claude Opus 4.6 on the SWE Bench Pro coding benchmark.
The family spans every major modality. MAI-Code-1 is a coding model already wired into GitHub Copilot and VS Code, while MAI-Image-2.5 (plus a faster Flash variant) handles text-to-image generation in PowerPoint and Microsoft Foundry. On the audio side, MAI-Voice-2 covers more than 15 languages with fine-grained control and MAI-Transcribe-1.5 claims state-of-the-art accuracy across 43 languages.
The models are rolling out across Microsoft Foundry, Copilot, VS Code, PowerPoint and OneDrive, alongside a new MAI Playground for developers and broader third-party access via providers such as Fireworks AI, Baseten and OpenRouter. The move signals Microsoft increasingly wants to own the AI stack it ships to enterprises, not just resell someone else's.