In a major strategic pivot, Microsoft has officially launched three new in-house foundational AI models, signaling its intent to reduce its heavy reliance on its partner, OpenAI. Announced by Mustafa Suleyman, CEO of Microsoft AI, these models—branded under the MAI (Microsoft AI) name—are designed for high-speed, enterprise-grade multimodal tasks and are available immediately on Microsoft Foundry (formerly Azure AI Studio) and the MAI Playground.
The trio aims to undercut competitors on price while delivering “best-in-class” performance in transcription, voice synthesis, and image generation.
1. MAI-Transcribe-1: The New Gold Standard for Speech
This speech-to-text model is being positioned as the world’s most accurate transcription engine for real-world, “messy” environments.
- Performance: Achieved the lowest average Word Error Rate (WER) of 3.8% on the FLEURS benchmark, beating OpenAI’s Whisper-large-v3 and Google’s Gemini 3.1 Flash across 25 major languages (including Hindi).
- Speed: Transcribes batch audio 2.5 times faster than Microsoft’s previous Azure Fast offering.
- Pricing: Aggressively priced starting at $0.36 per hour of audio.
- Use Case: Optimized for high-stakes environments like legal workflows, contact centers, and multilingual compliance-heavy industries.
2. MAI-Voice-1: High-Fidelity Speech Synthesis
MAI-Voice-1 is a text-to-speech model built to generate natural, expressive human voices that maintain emotional nuance even in long-form content.
- Lightning Speed: Capable of generating 60 seconds of audio in just one second.
- Custom Voices: Features a “Zero-Shot” capability, allowing developers to create secure, custom voice clones from just a few seconds of input audio.
- Pricing: Starts at $22 per 1 million characters.
- Integration: Already powering Copilot Audio Expressions and the new Copilot Podcasts feature.
3. MAI-Image-2: Turbocharged Visual Content
A direct successor to Microsoft’s previous image tools, MAI-Image-2 focuses on speed and “photorealistic” accuracy for professional layouts.
- Benchmark Success: Debuted as a top-three model family on the Arena.ai leaderboard.
- Speed Boost: Delivers at least twice the generation speed of its predecessor on Copilot and Foundry without sacrificing quality.
- Detail-Oriented: Specifically trained for “clear in-image text,” accurate skin tones, and natural lighting—addressing common pain points in AI-generated graphics and diagrams.
- Pricing: $5 per 1M tokens for text input and $33 per 1M tokens for image output.
4. Comparison: Microsoft MAI vs. Competitors
| Model Category | MAI Model | Key Advantage | Primary Rival |
| Transcription | MAI-Transcribe-1 | Lowest WER (3.8%) in 25 languages | OpenAI Whisper |
| Voice/Speech | MAI-Voice-1 | 60:1 generation speed ratio | ElevenLabs Scribe |
| Image | MAI-Image-2 | 2x speed; accurate in-image text | DALL-E 3 / Midjourney |
5. Why the “Foundry” Rebrand?
The release coincides with the rebranding of Azure AI Studio to Microsoft Foundry. This new unified platform is designed as a “one-stop-shop” for developers to build, red-team, and scale agentic AI applications.
By owning the models, the platform, and the hardware (via its custom Maia AI chips), Microsoft is attempting to “vertically integrate” its AI stack, much like Apple does with its hardware and software.
