At its Build 2026 developer conference, Microsoft showed off a new family of AI models built in-house. At the center are MAI-Code-1-Flash, a coding-focused model that turns written instructions into software code, and MAI-Thinking-1, a reasoning model that works through problems step by step. The company, which has long leaned on models from OpenAI and others, made clear its intent to handle core capabilities with models of its own.
MAI-Code-1-Flash, turning words into code
MAI-Code-1-Flash is Microsoft's first in-house coding model. It takes a request written in plain language and produces source code for apps and websites. So-called "vibe coding," where even people without a technical background can assemble software simply by describing what they want, has taken off lately, and Microsoft is aiming to capture that demand.
On the design side, the model uses a Mixture-of-Experts (MoE) structure that activates only the relevant experts for a given task. It holds 137 billion parameters in total, but only 5 billion are actually active in a single pass, balancing scale against light-footed operation. It can handle a context window of up to 256,000 tokens.
Its delivery is just as practical. The model is built into the GitHub Copilot code-completion service and the Visual Studio Code editor, rolling out to individual users in stages. Microsoft is also leading with low pricing, publishing rates of 0.75 USD (about 120 yen) per million input tokens and 4.50 USD (about 720 yen) per million output tokens. ※1 USD = 160 JPY
MAI-Thinking-1, a reasoning model built for efficiency
The other model, MAI-Thinking-1, reasons through a problem before answering. Here too the company stresses efficiency, delivering high performance at low token consumption. Tokens are the smallest units of data a model reads and writes, and how many are used translates directly into the cost developers pay.
It is a trillion-parameter-class model, but only 35 billion parameters run in a single pass, again a design that narrows the active portion. It will first be offered as a limited preview through Microsoft Foundry, a service for embedding models into applications, with customers invited to express interest in trying it before broad availability. Users can further raise the reasoning model's accuracy by feeding in their own data.
On training data, Microsoft says it built the model from scratch on clean, appropriately licensed data, without distilling outputs from third-party models. Mustafa Suleyman, head of Microsoft AI, said that in a case where the model was tuned for the consulting firm McKinsey, it beat OpenAI's GPT-5.5 while being 10 times more cost efficient.
The aim: thinning out reliance on OpenAI
Behind this announcement lies a shift in Microsoft's position within AI. The company has provided cloud infrastructure while investing 13 billion USD (about 2 trillion yen) in OpenAI and 5 billion USD (about 800 billion yen) in Anthropic, offering both companies' models through its own Azure cloud.
Yet the cost of using frontier models keeps rising. With models of its own, Microsoft can run them on Azure, avoid payments to outside parties, and pass the savings on to developers. Google moved first in this direction, announcing in May a coding-capable model, Gemini 3.5 Flash, that runs in its own data centers, and the push to keep costs down with in-house models is gathering pace across the industry.
At the same time, the very firms Microsoft has backed, OpenAI and Anthropic, are both moving toward going public. Anthropic filed for an IPO on June 1, and OpenAI is said to have an offering within the year in view. Against this backdrop, Microsoft is shifting its footing from being a user of models to being a maker of them, stepping into the foundations of AI itself. CEO Satya Nadella said from the stage, "What you just saw is a pretty significant shift. We believe the time has come for every company to move from consuming a frontier model to fully participating at the frontier in the frontier ecosystem."
A model family reaching into voice and images
What was shown goes beyond coding and reasoning. Microsoft also refreshed cloud models for speech recognition, synthetic voice generation, and image generation, and presented small models that run on Windows PCs. By lining up its own models for each use case, the company signals a stance of handling a broad range of AI layers itself.
Summary
At Build 2026, Microsoft unveiled an in-house family of models anchored by MAI-Code-1-Flash for coding and MAI-Thinking-1 for reasoning. Both follow an efficiency-first design that narrows the active parameter count, with low cost as a selling point. The move thins out the company's reliance on OpenAI and others it has invested heavily in, and marks a strategic turn toward joining the frontier as a model maker in its own right.
