Microsoft AI releases MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2 in Foundry

Background

Microsoft AI had already released its first in-house models before this announcement. In April 2026 the company made three of its models available to outside developers, not only for use inside its own products.^[1]

What happened

On 2 April 2026 Mustafa Suleyman announced three MAI models for developers: MAI-Transcribe-1, MAI-Voice-1 and MAI-Image-2.^[1] They became available in Microsoft Foundry, the company's platform for building with AI models, and in the MAI Playground, where people can try the models directly.^[1]

The three models cover three different tasks. MAI-Transcribe-1 turns speech into text. MAI-Voice-1 generates speech. MAI-Image-2 generates images from text.^[1][2]

Key details

MAI-Transcribe-1 covers 25 languages and runs 2.5 times faster than Azure Fast transcription. Pricing starts at $0.36 per hour of audio.^[1][2]

MAI-Voice-1 supports custom voices built from a few seconds of audio, priced from $22 per million characters. Custom voices require an approval process before they can be used.^[1][3] MAI-Image-2 debuted as a top-3 model family on the Arena.ai leaderboard. It is priced from $5 per million text input tokens and $33 per million image output tokens.^[1][2] Microsoft said the model was rolling out across Bing, PowerPoint and Copilot.^[1]

Impact

Microsoft said the three models had been red-teamed before release, and that it had published model cards for each one.^[1] A model card is a short document that explains what a model is for and what its limits are.

For the field, the announcement marked a shift in how Microsoft presents its own models. They were no longer only features inside Microsoft products. They became offerings that developers could build on, with stated prices and published documentation, next to models from other companies on the same platform.^[1][3]

Developers usually pay for cloud AI models by usage rather than by a flat fee. For text, prices are often quoted per million tokens, where a token is a short piece of a word. For speech, the price is quoted per hour of audio, and for voices, per million characters. The three prices in this announcement follow those conventions, so developers can estimate costs for their own workloads.^[1][2]