Microsoft Launches MAI Models to Challenge OpenAI Directly — Built by Teams of Under 10 Engineers

April 2, 2026 · 8 min read · by Connie

For the past three years, Microsoft's AI strategy has been synonymous with one name: OpenAI. That is changing. On April 2, 2026, Microsoft released its own foundational AI models under the MAI brand — speech, voice, and images — developed in-house by Mustafa Suleyman's Superintelligence team. The models are immediately available through Microsoft Foundry and are priced to undercut competitors.

The Three MAI Models

MAI-Transcribe-1: Speech-to-Text

MAI-Transcribe-1 is Microsoft's answer to OpenAI's Whisper. It achieves the lowest average Word Error Rate (WER) on the FLEURS benchmark across the 25 languages most commonly used in Microsoft's products — averaging just 3.8% WER. It outperforms OpenAI's Whisper-large-v3 across all 25 languages and Google's Gemini 3.1 Flash on 22 of 25 languages.

Speed is a headline advantage: MAI-Transcribe-1 is 2.5× faster than Microsoft's existing Azure Fast transcription offering. It is optimized for real-world noisy environments — call centers, conference rooms, industrial settings — rather than clean studio audio. Microsoft is testing integrations with Copilot and Teams. Pricing: $0.36 per hour.

MAI-Voice-1: Text-to-Speech

MAI-Voice-1 generates 60 seconds of natural-sounding audio in under one second on a single GPU. The model supports custom voice creation from short audio snippets — useful for enterprises that want a consistent brand voice without extensive recording sessions.

Pricing at $22 per million characters positions it as a competitive alternative to OpenAI's TTS models. The sub-one-second generation time at 60 seconds of audio is the metric Microsoft is leading with — relevant for real-time applications like live translation and voice interfaces.

MAI-Image-2: Text-to-Image

MAI-Image-2 ranks in the top three on the Arena.ai text-to-image leaderboard and delivers at least 2× faster generation times than its predecessor. Microsoft is rolling it out across Bing and PowerPoint as an embedded capability. Pricing: $5 per million tokens for text input, $33 per million tokens for image output. Microsoft claims it can run on approximately half the GPU footprint of comparable competitor models.

MAI Models vs. OpenAI and Google: Benchmark Comparison

Why Microsoft Is Building Its Own Models Now

The pivot to in-house model development is the direct result of a renegotiated contract with OpenAI finalized in October 2025. The original partnership terms prohibited Microsoft from independently pursuing artificial general intelligence — a restriction that effectively barred in-house frontier model development. The revised agreement removed that constraint while preserving Microsoft's license rights to OpenAI's technology through 2032.

Mustafa Suleyman, Microsoft's CEO of AI, handed off day-to-day Copilot engineering to Jacob Andreou in March 2026 and shifted to leading the Microsoft AI Superintelligence team full-time. Suleyman has been planning this transition for nearly nine months, developing the technical foundation for Microsoft to build "humanist superintelligence" — systems that are intellectually advanced but strictly aligned with human interests.

The financial pressure is real. Microsoft's stock fell roughly 17% year-to-date in early 2026 as investors demanded proof that massive AI infrastructure spending would yield returns. In-house models that run on half the GPU footprint of competitors directly reduce per-query costs for products like Copilot — translating infrastructure investment into margin improvement.

Microsoft Foundry: The New AI Distribution Platform

Microsoft Foundry is the platform where all three MAI models are available. It also hosts OpenAI and Anthropic models, positioning Microsoft as a multi-model distribution layer rather than an exclusive OpenAI reseller. Developers can access MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 immediately through Foundry, alongside the MAI Playground for evaluation.

This is the same strategic move that Microsoft Copilot's Critique mode illustrated at the application layer: Microsoft is positioning itself as the platform where multiple AI providers compete, rather than being locked into any single model vendor.

Pricing: How MAI Models Stack Up

What This Means for Developers

The MAI launch signals a meaningful expansion of available high-quality speech and image models. MAI-Transcribe-1's performance across 25 languages — including non-English languages where Whisper has historically underperformed — is particularly relevant for multilingual enterprise deployments.

For developers already building on AI coding tools or integrating speech features into products, MAI-Transcribe-1 and MAI-Voice-1 are now credible alternatives to evaluate against Whisper and Google Speech-to-Text. The infrastructure advantage — half the GPU footprint — translates to lower latency and cost at scale.

The broader implication: the "use OpenAI for everything" default is no longer the obvious choice. Microsoft has now joined Anthropic, Google, and Alibaba in offering competitive foundation model APIs across multiple modalities. See our full model comparison for a broader framework on choosing the right model for your use case.