By Microsoft just released three AI models built entirely in-house, and none of them are large language models. That's the interesting part.
MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 launched Wednesday through Microsoft Foundry and a new MAI Playground. They handle speech-to-text, voice generation, and image creation, respectively. All three came out of the MAI Superintelligence team that Mustafa Suleyman formed in November 2025, just months after Microsoft and OpenAI renegotiated their partnership in a way that, according to Suleyman, "unlocked" Microsoft's ability to pursue this kind of work independently.
Think of it like a restaurant chain that's been buying all its ingredients from one premium supplier. The food is great, but the supplier sets the prices, controls the menu, and could theoretically open competing locations. So the chain quietly starts growing its own produce. Not everything at once. Just the ingredients where they can match or beat the quality, at a fraction of the cost.
That's what Microsoft is doing with AI models, and the early numbers suggest they're not bluffing.
The transcription model is the headliner
MAI-Transcribe-1 is the one Suleyman is most vocal about, and the benchmarks explain why. According to Microsoft, the model achieves the lowest average Word Error Rate on the FLEURS benchmark across the top 25 languages by Microsoft product usage, averaging 3.8% WER. It reportedly beats OpenAI's Whisper-large-v3 on all 25 languages, Google's Gemini 3.1 Flash on 22 of 25, and ElevenLabs' Scribe v2 and OpenAI's GPT-Transcribe on 15 of 25 each.
Batch transcription runs 2.5 times faster than Microsoft's own Azure Fast offering. Suleyman told VentureBeat the model operates at "half the GPU cost of state-of-the-art competition," and it was built by a team of just 10 people.
Microsoft is already testing it inside Copilot's Voice mode and Microsoft Teams for conversation transcription, which tells you something about how quickly the company intends to swap out third-party dependencies with its own models.
MAI-Voice-1, the voice generation model, can produce 60 seconds of audio in one second and supports custom voice creation. MAI-Image-2, which first appeared on MAI Playground on March 19, handles image generation and is now available through Foundry alongside the other two.
Pricing is aggressive: MAI-Transcribe-1 starts at $0.36 per hour, MAI-Voice-1 at $22 per million characters, and MAI-Image-2 at $5 per million tokens for text input and $33 per million tokens for image output.
Why the modality choices matter
Microsoft didn't start with a large language model. That's deliberate. Building a competitive LLM from scratch would take years and directly compete with OpenAI, a company Microsoft has invested more than $13 billion in. Starting with transcription, voice, and image generation lets Microsoft build real model-development muscle in areas where the competitive dynamics are less fraught.
But the long-term trajectory is clear. Suleyman has said he wants to be "completely independent" if needed, and the MAI Superintelligence team's mandate extends well beyond multimodal utilities. These three models are proof-of-concept, not the endgame.
This also arrives at a tense financial moment. Microsoft's stock just closed its worst quarter since the 2008 financial crisis, with investors questioning whether hundreds of billions in AI infrastructure spending will pay off. Models that reduce Microsoft's own cost of goods sold, running inside products that already have hundreds of millions of users, are one way to answer that question without waiting for a speculative frontier LLM to materialize.
The broader pattern across the industry is hard to miss. Mistral is building its own data center in Europe to own its compute stack. Every major player is moving toward vertical integration, controlling more of the pipeline from silicon to model to product. Microsoft is following the same playbook it uses with chips: build your own while still buying from outside.
What we don't know yet
- Whether these benchmark results hold up under independent third-party testing, or if they reflect cherry-picked conditions favorable to Microsoft's models.
- How quickly the MAI Superintelligence team will move from multimodal utilities to a frontier LLM, and what that does to the OpenAI relationship when it happens.
- Whether the "team of 10 people" framing reflects the actual resource allocation, or if shared Microsoft infrastructure and data are doing heavy lifting that a headcount number doesn't capture.
What comes next
Suleyman wrote in the announcement blog post that users should expect "more models from us soon in Foundry and directly in Microsoft products and experiences." Features like diarization, contextual biasing, and streaming for the transcription model are listed as coming soon.
The real question is when "more models" includes a large language model. Suleyman has the team, the compute, the contractual freedom, and now the first shipped products to point to as evidence his group can execute. If these three models deliver on their cost and quality promises at scale, the case for building a Microsoft-native LLM gets harder to argue against, both inside the company and out.
For now, Microsoft is still OpenAI's biggest customer and biggest investor. But it's also, quietly, becoming a competitor.
Marcus Webb covers AI for The Daily Vibe.
