At Build 2026, Microsoft announced seven new AI models developed in-house, including its first reasoning model. The company also introduced a new tuning method and an autonomous background agent.
The centerpiece is MAI-Thinking-1, Microsoft's first reasoning model. According to Microsoft AI chief Mustafa Suleyman, it's a 1-trillion-parameter model with 35 billion active parameters and a 128,000-token context window, built for multi-step instructions, long contexts, and code generation.
Microsoft says MAI-Thinking-1 matches leading models on key software engineering benchmarks and was preferred over Anthropic's Sonnet 4.6 in internal blind comparisons. The model was trained from scratch on clean data without distillation from third-party models, according to Suleyman. That's a not-so-subtle jab at practices at other labs. A look at the published benchmarks, though, puts the model roughly on par with Deepseek V3.2.
A model family spanning six task areas
Beyond the reasoning model, the MAI family includes six more systems. MAI-Code-1-Flash is an agentic coding model with 5 billion parameters that Microsoft says is comparable to Anthropic's Haiku but cheaper to run. It's integrated into GitHub Copilot and Visual Studio Code.
