Microsoft has launched three new in-house AI models through its Foundry platform — MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 — marking the company's clearest move yet to reduce its reliance on OpenAI for foundational AI capabilities.

What's in the bundle

MAI-Transcribe-1 is a speech-to-text model priced at $0.36 per hour of audio. Microsoft says it outperforms Whisper-large-v3 across all of the top 25 languages by Microsoft product usage, and beats Gemini 3.1 Flash on 11 of the 14 remaining. It's now in public preview on Foundry and already powering Copilot's Voice Mode and Teams transcription.

MAI-Voice-1 handles text-to-speech and can generate 60 seconds of audio in just one second. It's priced at $22 per 1M characters and currently drives Copilot's Audio Expressions and Podcast features.

MAI-Image-2 is the image generation model, with improved lighting, skin tone rendering, and text fidelity compared to its predecessor. Pricing starts at $5 per 1M text input tokens and $33 per 1M image output tokens.

Why it matters

All three models are built by Microsoft's internal MAI Superintelligence team — not licensed from OpenAI. The company frames them as "Humanist AI," optimized for how people actually communicate. For developers, the full stack is available today via Microsoft Foundry; the MAI Playground offers a public demo (US only).

The move signals that Microsoft is building real independence in its AI stack, and is willing to compete directly with its biggest partner on model quality and price.