Summary: Microsoft AI CEO Mustafa Suleyman said on April 2, 2026 that MAI-Transcribe-1, MAI-Voice-1, and MAI-Image-2 entered public preview in Foundry and MAI Playground. Microsoft cites 25-language transcription accuracy, sub-second generation of 60 seconds of audio, third-place Arena.ai image ranking, and in-image text rendering. Pricing starts at $0.36 per transcription hour, $22 per million voice characters, and $5 per million text-input tokens and $33 per million output tokens for image generation.
On April 2, 2026, Microsoft AI CEO Mustafa Suleyman confirmed that three models built inside Microsoft — MAI-Transcribe-1 for turning speech into text, MAI-Voice-1 for synthesizing speech, and MAI-Image-2 for generating images — have entered public preview through Microsoft Foundry and MAI Playground. With this release, Microsoft’s first-party AI lineup stops being text-only: audio and vision are now covered, and the pitch is aimed at enterprises via APIs rather than at consumers through demos.
Suleyman’s angle deserves attention: these models were put together from scratch with the enterprise in mind, and the announcement stresses that one set of weights serves both Microsoft’s own products and the capacity developers rent by the hour. For organizations already inside the Azure ecosystem, that folds two separate vendor questions — can it do the job, and does it clear compliance — into a single decision.
Three New Arrivals in Foundry
Microsoft’s Tech Community post describes Foundry as “the most complete AI and app agent factory,” and the trio of MAI models are its latest additions. MAI-Transcribe-1 and MAI-Voice-1 run on Azure AI Speech resources you already have, so no pipeline has to be rebuilt; MAI-Image-2 sits in Foundry’s standard model catalog. These are also the models behind Copilot, Bing Image Creator, PowerPoint, and Azure Speech — Microsoft’s own products and the external API draw on identical weights.
Speech In, Speech Out: Microsoft’s Own Audio Layer
The headline claim for MAI-Transcribe-1 is accuracy across 25 languages: on the FLEURS benchmark it takes first place in 11 of the top 25 languages, outscores Whisper-large-v3 in the other 14, and outperforms Gemini 3.1 Flash in 11 of those 14. On cost, Microsoft says GPU spend runs about 50% below “leading alternatives,” batch transcription runs 2.5x faster than Azure Fast, and the rate starts at $0.36 per hour. The model is not untested, either: it already does the transcription and dictation work inside Copilot’s Voice Mode, so it has been carrying production traffic at Copilot scale ahead of any external launch. Microsoft also states plainly that the model carries documented human biases from its training data and explains how it handles them — an admission you seldom find in a launch post, and a relevant one when transcripts feed automated decisions downstream.
MAI-Voice-1 is aimed at long-form material, holding onto emotional range and a consistent speaker identity throughout; Microsoft’s own description says writing speech with it is “meant to feel more like drawing than dictating.” The performance figure: one GPU turns out 60 seconds of audio in under a second, and the price is $22 per million characters. Personal Voice can replicate a person’s voice from a 10-second sample, but every use goes through an approval process — friction built in on purpose. The existing gallery of 700+ voices in Azure Speech stays available next to it.
MAI-Image-2: Landing at Number Three Out of the Gate
In its first appearance on the Arena.ai image leaderboard, MAI-Image-2 took the number-three spot; Microsoft adds that the family holds a firm place in the top three and runs at least twice as fast as the previous generation. The announcement spends much of its energy on in-image text rendering — producing readable text inside a generated image, long a sore point for diffusion models. WPP is on board as an early enterprise adopter, and its global Chief Creative Officer Rob Reilly called the model “a genuine game-changer” for campaign-ready work. Copilot, Bing, and PowerPoint will adopt it in phases, so the image experience consumers meet across Microsoft’s products is set to run on the same engine developers can reach from Foundry.
The Price Tags and the Responsible AI Caveats
The numbers on the label: transcription at $0.36 per hour, voice at $22 per million characters, and image generation at $5 per million tokens of text input with output at $33 per million tokens. Measured against the wider API market, Microsoft is plainly leaning on its first-party infrastructure to drive unit pricing down — the same battleground as the model-economics war we mapped in our Replit free-mode analysis: the player who pushes the marginal cost of intelligence lower takes the distribution. The caveats are spelled out too: cloning a voice requires approval, and the image and voice models arrive with guardrails built in. Responsible AI limits are part of the product’s design from the start, not an add-on.
The Takeaways for Developer and Product Teams
Three consequences worth acting on. One, audio pipelines get another legitimate supplier: transcription and voice generation draw on a single Azure Speech resource, so switching over stays cheap. Two, first-party products running on the same models tells you the usage scale is genuine — every day of Copilot traffic doubles as a stress test for these weights. Three, the image API’s output price ($33 per million tokens) earns a row in your comparison sheet, above all for marketing workloads that lean on heavy in-image typography — the old failure mode that sent work back through an editing tool for a second pass, now handled by the model itself.