Microsoft Launches Three In-House AI Models in Challenge to OpenAI

On April 2, Mustafa Suleyman, CEO of Microsoft AI, posted on X: "Today we're announcing 3 new world class MAI models, available in Foundry." The three models are MAI-Transcribe-1 for speech-to-text, MAI-Voice-1 for voice synthesis, and MAI-Image-2 for image generation. All three are available through Microsoft Foundry, the company's enterprise developer platform.

Suleyman described the approach behind the models in a blog post: "At Microsoft AI, we're building Humanist AI. We have a distinct view when creating our AI models, putting humans at the center, optimizing for how people actually communicate, training for practical use." He told VentureBeat that MAI-Transcribe-1 delivers transcription capability at "half the GPUs of the state-of-the-art competition."

MAI-Transcribe-1 supports transcription across 25 languages and runs at 2.5 times the speed of Microsoft's existing Azure Fast transcription service. It now powers Copilot's Voice Mode, replacing the previous transcription layer. At launch it supports batch transcription of pre-recorded audio rather than real-time streaming. MAI-Voice-1 generates 60 seconds of audio per second of processing time and supports custom voice creation. MAI-Image-2 ranks in the top three on the Arena.ai image generation leaderboard and is rolling out across Bing and PowerPoint.

The OpenAI Relationship, Reread

Microsoft invested $13 billion into OpenAI across multiple rounds. In return it got the right to deploy OpenAI models through Azure and a share of profits above a cap. That arrangement looked straightforward when GPT-4 was the dominant model. It is more complicated now that Google's Gemini models are strong, Anthropic's Claude is a serious enterprise option, and Microsoft is shipping models it built in-house.

TechCrunch described the release as Microsoft "taking on AI rivals." The MAI prefix is explicit branding. MAI stands for Microsoft AI, and labelling these models under that brand rather than integrating them invisibly into Copilot is a positioning decision. The MAI Superintelligence team, which built these models, was formed and announced in November 2025.

The OpenAI partnership remains active. OpenAI models are still available through Azure and the commercial relationship generates revenue for both companies. The MAI launch makes explicit what has been building for months: Microsoft wants to be a model developer, not only a distribution platform for others.

What This Changes for Developers

Developers building on Microsoft Foundry now have Microsoft's own models as a native option alongside OpenAI and third-party alternatives. The practical effect is optionality: an enterprise can run a Microsoft AI stack end-to-end without touching OpenAI's API. Latency, pricing, and quality in production conditions will determine how many do.

GeekWire said the models represent Microsoft's clearest move yet to expand beyond its reliance on OpenAI's technology. The batch-only constraint on MAI-Transcribe-1 at launch is a genuine limitation for real-time applications, but the model's speed and language coverage make it competitive for the document and media processing workloads that account for most enterprise transcription volume.

The Bigger Picture

OpenAI's $852 billion valuation is built on ChatGPT's consumer reach and API revenue. A large portion of that API revenue flows through enterprise agreements, the same enterprises Microsoft is now offering native alternatives to. Enterprise AI procurement cycles run 18 months. The MAI models will not displace OpenAI overnight. But the negotiating dynamic between Microsoft, OpenAI, and enterprise buyers has changed with the announcement.