Microsoft's first reasoning model is one of 7 AIs just released at Build - what we know so far
Microsoft unveiled seven new AI models at Build, including its first reasoning model, plus new coding, image, voice, and transcription systems.
Intelligence analysis by GPT-5.4 Mini
Microsoft used Build to show a broader AI stack, led by MAI-Thinking-1, its first reasoning model, and MAI-Code-1 for coding. The company also introduced new image, transcription, and voice models, and said it is working with Mayo Clinic on a healthcare model.
Microsoft added a bunch of new AI helpers, like a thinker, a coder, a picture maker, and a voice listener. It is like stocking a toolbox with different tools instead of one big hammer, so each job can be done better.
Analysis
What Microsoft announced
Microsoft AI introduced seven new models at Build, with MAI-Thinking-1 as its first reasoning model. The company says it was trained on “enterprise-grade, clean and commercially licensed data,” and that independent reviewers preferred it over Anthropic's Sonnet 4.61 in a blind test. Microsoft also says the 35-billion-parameter model matches Anthropic Opus 4.6 on SWE Bench Pro for coding.
The rest of the lineup
MAI-Code-1 is the coding model, described as “ultra-efficient” and tuned for GitHub. It is coming to Copilot and VS Code immediately. MAI-Image-2.5 and its flash version are Microsoft’s first text-to-image and image-to-image models; the company says they outperformed Nano Banana Pro on ELO and were already showing strong placement on LM Arena.
Microsoft also launched MAI-Transcribe-1.5, which it says offers state-of-the-art accuracy across 43 languages, with streaming support coming soon. MAI-Voice-2 and its flash companion extend language coverage by 15 more languages than MAI-Voice-1.
Where they will ship
The image models are live in PowerPoint and Foundry, and are rolling out in OneDrive. MAI-Thinking-1 is available in Microsoft Foundry in private preview. The full set will also be available on Fireworks AI, which Microsoft said is now generally available on Foundry, plus Baseten and Open Router.
Security and health
Microsoft emphasized that “everything is watermarked from scratch” and said some models offer cost efficiency gains of up to 10x versus comparable competitors. Suleyman closed by announcing a Mayo Clinic collaboration to build a frontier healthcare model, while noting that privacy, security, and hallucinations remain concerns in medical AI.
Key points
- Microsoft announced seven new AI models at Build, including its first reasoning model, MAI-Thinking-1.
- The company says MAI-Thinking-1 was trained on commercially licensed data and performed well in blind tests and coding benchmarks.
- MAI-Code-1 is aimed at GitHub and is shipping to Copilot and VS Code.
- Microsoft launched new image, transcription, and voice models, with some already live in PowerPoint and Foundry.
- The company also announced a Mayo Clinic partnership to build a healthcare-focused frontier model.
If these models perform as advertised, Microsoft can give its products more specialized AI tools for writing code, making images, transcribing speech, and handling multi-step tasks. Wider availability through Foundry, Copilot, VS Code, PowerPoint, and partner platforms could also make adoption faster for developers and businesses.
The models are still new, and some are only in preview, so real-world performance may fall short of the benchmark claims. The healthcare push also faces familiar risks around privacy, security, and hallucinations, which the article notes remain concerns.
