Microsoft is reportedly pushing deeper into foundation model development with an internal language model family known as MAI. The work is being led under Mustafa Suleyman, Microsoft’s "CEO of AI," and is aimed at reaching capabilities now associated with OpenAI and Anthropic.
The reported plan is not only to use these models inside Microsoft. The company could also open API access to outside developers before year’s end, a move that would place Microsoft more directly in the market for AI model services.
What Microsoft is building with MAI
According to The Information, Microsoft’s AI team has reached an important milestone with MAI. Testing indicates that the models perform at nearly the same level as leading offerings from OpenAI and Anthropic on standard benchmarks.
The MAI family also includes a reasoning model built specifically to match OpenAI’s o1 capabilities. That detail matters because reasoning models are designed for tasks that need more deliberate, step-by-step problem solving rather than only fast text generation.
The source describes these models as much larger than Microsoft’s earlier Phi series. Phi focused more on balancing cost and performance, while MAI appears to be a more ambitious attempt to compete with frontier model providers.
Microsoft has not been building in only one direction. The company is also evaluating models from xAI, Meta, and DeepSeek in Copilot as possible replacements for its current OpenAI models. That wider testing suggests Microsoft wants more flexibility in how it powers its AI products.
Why API access would change the stakes
If Microsoft opens MAI through an API later this year, the company would move beyond being a major AI product distributor and cloud partner. It would become a more direct supplier of models to developers, putting it closer to the position occupied by OpenAI and other AI labs offering similar services.
For outside developers, API access would be the practical turning point. It would mean MAI is not only an internal Microsoft project, but a model family that can be tested, integrated, and compared in real applications.
The source does not specify pricing, availability terms, or exact launch timing beyond the possibility of API access before year’s end. It also does not say which MAI models would be made available first. What it does show is that Microsoft is preparing the pieces needed for a more independent AI platform strategy.
- MAI is Microsoft’s reported internal language model family.
- API access could arrive before year’s end.
- Copilot is already being used to evaluate models from xAI, Meta, and DeepSeek.
- OpenAI and Anthropic are the capability benchmarks Microsoft is reportedly trying to match.
A push for independence from OpenAI
Microsoft’s AI strategy has been closely tied to OpenAI, but the MAI effort points toward a desire for more control. The company could reduce its reliance on its AI partner by building strong internal models and by testing alternatives from other providers.
The path has not been simple. The Information reports that MAI’s year-long development included technical difficulties, changes in direction, and the departure of several key team members. Those challenges unfolded while OpenAI continued releasing new versions of its models.
One major departure was Sébastien Bubeck, who headed the Phi project. He left for OpenAI and took several Microsoft researchers with him, according to the source. That made the internal model effort harder at a time when Microsoft was trying to advance its own capabilities.
The report also says Suleyman’s frustration with OpenAI grew during this period, especially because OpenAI would not reveal details about the inner workings of o1. Even so, the AI team under Karén Simonyan managed to develop comparable reasoning abilities using chain-of-thought techniques.
How MAI connects to Microsoft’s earlier AI work
Reports from spring 2024 mentioned a Microsoft model called MAI-1 with approximately 500 billion parameters. Microsoft’s CTO Kevin Scott later confirmed MAI’s existence on LinkedIn, while also emphasizing that Microsoft’s partnership with OpenAI was continuing.
The new model also incorporates technology and training data from Suleyman’s AI startup Inflection, which Microsoft acquired for $650 million. That connection gives the MAI project a direct link to Microsoft’s broader effort to absorb AI talent, tools, and model-building experience.
The result is a strategy with several layers. Microsoft is building its own models, evaluating outside models, and maintaining its OpenAI relationship. Those moves are not identical, but they all support the same practical goal: giving Microsoft more options for the AI systems behind its products and services.
For the AI market, MAI is significant because Microsoft is not approaching the space as a new entrant with limited distribution. It already has Copilot, an existing OpenAI partnership, and the ability to offer APIs to developers if the models are ready. If the reported API plan moves ahead, MAI could become one of the clearest signs yet that Microsoft wants to compete in foundation models with its own technology, not only through partners.