Microsoft unveils seven in-house AI models to break free from OpenAI and go hunting in the lands of Anthropic and Google

For years, Microsoft's AI was mainly that of OpenAI, the creator of ChatGPT, in which the giant has invested billions. Practical, but uncomfortable: depending on a partner for the most strategic technology of the decade ends up itching. So Microsoft is changing gear. Under the leadership of Mustafa Suleyman, its AI boss and former co-founder of Google DeepMind, the company is launching its own family of models, called MAI. And the tone has changed. Here are the seven models:

- MAY Image-2.5

- MAY Image-2.5-Flash

- MAY Transcribe-1.5

- MAY Thinking-1

- MAY Voice-2

- MAY Voice-2-Flash

- MAY Code-1-Flash

The headliner is MAI-Thinking-1, presented as Microsoft's first in-house reasoning model, this type of AI which takes the time to break down a problem before responding. The detail that counts: it was trained “from scratch”, without distillation, that is to say without learning by copying the answers from another existing model. A way to claim a truly original AI, not a derivative. It is a medium-sized model, 35 billion active parameters, designed for a good performance-to-cost ratio rather than for excess.

Microsoft compares itself head-on to Claude and Google

This is where Microsoft flexes its muscles, and you have to take these numbers for what they are: its own. According to its own tests, independent evaluators, via a rating partner, blindly prefer MAI-Thinking-1 to Claude Sonnet 4.6 from Anthropic, and the model would equal Opus 4.6 on the code, measured on the SWE Bench Pro benchmark (we are today at Opus 4.8).

Microsoft is adding more on the image side with MAI-Image-2.5, which it claims to surpass Google's Nano Banana Pro in the ELO ranking. Statements to be handled with a grain of salt, since no independent comparison has yet validated them, but the symbol is strong: Microsoft is now attacking those who were its partners or suppliers.

The MAI family doesn’t stop there. Microsoft also announces MAI-Image-2.5 and its Flash variant, already integrated into PowerPoint, a transcription model covering 43 languages, synthetic voices in more than 15 additional languages, and MAI-Code-1, an ultra-efficient code model designed for GitHub and already present in Copilot and VS Code. Seven models in one go, covering text, image, voice and code. This is a complete range, not an isolated demo.

To go further

How to install a LLM ChatGPT model on PC or Mac locally? Here is the ultimate guide for everyone

The real bet: own your AI rather than rent it

Beyond the numbers, the philosophy is clear. Microsoft is banking on sovereignty: these models are trained on clean data and under commercial license, and the company insists that they run in a framework where your data does not feed a third party. This is the same common thread that we find throughout the current strategy: running AI as close as possible to you, locally when possible, rather than sending everything to someone else's cloud. Moreover, these models will also be accessible elsewhere than at Microsoft, on platforms like Fireworks AI, Baseten or OpenRouter.

It remains to be seen whether reality will follow the speech. Microsoft is not the first to promise in-house AI that rivals the best, and until independent tests pit MAI-Thinking-1 against Claude and Google's models, these rankings will remain keynote arguments. But there is no longer any doubt about the direction: after years of relying on OpenAI, Microsoft clearly wants to stand on its own two feet. And when the world's number one software company decides to manufacture its own intelligence, the entire balance of the sector begins to shift.