Microsoft Launches Proprietary AI Models to Reduce OpenAI Dependence
Microsoft announced two new AI models at its Build developer conference in San Francisco on Tuesday: MAI-Code-1-Flash, a coding model that generates source code from text descriptions, and MAI-Thinking-1, a reasoning model designed for cost efficiency. The moves signal Microsoft's shift from primarily supplying cloud infrastructure to competing directly with OpenAI, Anthropic, and Google in model development.
The economic incentive is straightforward. Microsoft has invested $13 billion in OpenAI and $5 billion in Anthropic while reselling their models through Azure. Building proprietary models lets Microsoft run inference on its own infrastructure and avoid paying third parties, with savings passed to developers as token costs-the unit that determines pricing-drop.
Cost Advantages in Coding
MAI-Code-1-Flash is available now through GitHub Copilot and Visual Studio Code. Microsoft describes it as "inference ultra-efficient," meaning it uses fewer tokens to produce results, reducing developer costs compared to competing coding models.
The reasoning model, MAI-Thinking-1, enters private preview through Microsoft Foundry. Developers can request access before broader availability. Kyle Daigle, Microsoft's developer marketing chief and GitHub operating chief, emphasized the model's "low-token cost" design.
Microsoft CEO Satya Nadella said the company achieved 10 times better cost efficiency than OpenAI's GPT-5.5 after refining models for consulting firm McKinsey. Customers can also improve accuracy by incorporating their own proprietary data into the reasoning model.
Broader Model Portfolio
Beyond coding and reasoning, Microsoft released updated cloud-based models for speech recognition, synthetic voice generation, and image generation. The company also introduced small Aion models designed to run on Windows PCs without cloud dependencies.
The announcements come as both OpenAI and Anthropic pursue public offerings. Anthropic confidentially filed for an IPO on June 1. Google released its own efficient coding model, Gemini 3.5 Flash, in May, running on its own data centers.
For development teams, the shift means more model options and potentially lower costs, but also more evaluation work to determine which models fit specific use cases and performance requirements.
Learn more: AI Coding Courses and Generative Code Courses can help developers understand how to work with these new tools effectively.
Your membership also unlocks:
