Microsoft Debuts In-House AI Models Built Without OpenAI

For years, Microsoft's AI strategy could be summed up in one word: OpenAI. The company poured more than $13 billion into the lab and wired its models into everything from Windows to Office. At Build 2026, Microsoft made its clearest move yet toward standing on its own.

The company unveiled two models built entirely in-house. MAI-Thinking-1 is its first reasoning model, and the detail Microsoft kept returning to is how it was trained: from scratch, on commercially licensed enterprise data, with no distillation from third-party systems. That is a pointed contrast with the common industry practice of bootstrapping a new model by learning from an existing one, often a competitor's. Microsoft is saying, in effect, that none of OpenAI's DNA is inside this thing.

The specifications

MAI-Thinking-1 is a sparse mixture-of-experts model with about 35 billion active parameters, roughly a trillion in total, and a 256,000-token context window. On AIME, the mathematics-and-reasoning benchmark, Microsoft reports scores of 97.0 percent for the 2025 set and 94.5 percent for 2026. On SWE-Bench Pro, a software-engineering test, the company says it matches Claude Opus 4.6, and that in blind side-by-side trials run by its independent rating partner Surge, evaluators preferred it to Claude Sonnet 4.6.

Those are Microsoft's own numbers, and the model is only in private preview through Microsoft Foundry for now, so the usual caveat applies until outside labs can check the work.

The second model is more immediately useful to most people. MAI-Code-1 is a coding model tuned for the kind of work GitHub sees all day, and it is already shipping inside GitHub Copilot and Visual Studio Code. The lighter MAI-Code-1-Flash is rolling out to about 10 percent of individual users, and anyone who picks "Auto" in the VS Code model selector may quietly find themselves routed to it. Both models are also available through Fireworks AI, Baseten and OpenRouter.

Why it matters

The strategic logic is straightforward. Running someone else's frontier model at Microsoft's scale is enormously expensive, and depending on a single partner for a core capability is a risk no company that size is comfortable carrying. Owning the model means controlling the cost and the roadmap. Microsoft framed the launch as a way to lessen its reliance on OpenAI and lower prices for developers.

It also fits a pattern we have been tracking. The action in AI right now is less about ever-bigger models and more about efficient, purpose-built ones, a shift we wrote about in why a bigger model is not always a smarter one. A 35-billion-active-parameter model that trades blows with the largest systems on coding is exactly the kind of result that argument predicted.