Microsoft Builds Its Own AI Models to Cut Costs and Challenge OpenAI

Microsoft has long been the backbone of the AI boom, supplying cloud infrastructure and pouring billions into OpenAI and Anthropic. But at its Build developer conference in San Francisco, the company signaled a shift: it’s now building its own models to compete directly with those partners.

The headline release is MAI-Code-1-Flash, Microsoft’s first model that turns plain-language descriptions into working source code for apps and websites. This “vibe coding” market is exploding, letting everyone from seasoned developers to hobbyists generate software with text prompts. By running MAI-Code-1-Flash on its own Azure servers, Microsoft avoids paying third-party providers like OpenAI—savings it can pass to developers as costs for leading models climb.

Alongside the coding model, Microsoft introduced MAI-Thinking-1, a reasoning model built for efficiency. “It’s medium-sized, high-performance, and low-token cost,” wrote Kyle Daigle, Microsoft’s head of developer marketing and GitHub operations. Tokens are how developers pay for model usage, so lower costs matter. The model is available in private preview through Microsoft Foundry, and customers can boost its accuracy by feeding in their own data.

Microsoft is hedging its bets as OpenAI and Anthropic eye public listings. Anthropic filed confidentially for an IPO Monday; OpenAI may follow this year. Microsoft holds $13 billion in OpenAI and $5 billion in Anthropic, but CEO Satya Nadella sees a new era: “Every company should move from consuming frontier models to participating in the frontier ecosystem.”

Early tests with consulting firm McKinsey showed Microsoft’s reasoning model outperformed OpenAI’s GPT-5 with tenfold better cost efficiency, according to Microsoft AI CEO Mustafa Suleyman. The coding model is now live in GitHub Copilot and Visual Studio Code. Microsoft also unveiled updated cloud models for speech, voice synthesis, image generation, and small AI models that run on Windows PCs.