Microsoft has been paying OpenAI's bills for years. It invested over $13 billion since 2019 and, in return, got to power its products with OpenAI's models. Every time someone used Copilot in Word, or got an AI-generated answer in Bing, OpenAI's technology was doing the work. That arrangement is now changing.
At its Build 2026 developer conference on June 2, Microsoft launched MAI-Thinking-1, its first reasoning AI model. A reasoning model is one that takes extra time to think through a problem step by step before answering, rather than producing an immediate response. This matters for complex tasks: legal analysis, contract review, financial modeling, multi-step planning. Microsoft says the model was built entirely from its own data and training process, without copying or learning from any other company's AI output.
The model has 35 billion parameters, which is mid-sized by industry standards. Microsoft claims it performs at the level of leading models on software engineering tasks and matches Anthropic's Claude Sonnet 4.6 in preference tests where users compare answers side by side without knowing which model produced them. These are company-reported numbers, so independent verification will take time.
MAI-Thinking-1 was not Microsoft's first in-house AI model overall. The company began testing a basic text model called MAI-1-preview in August 2025, and released three specialized models in April 2026 covering speech-to-text, voice generation, and image creation. But MAI-Thinking-1 is the first designed for the kind of complex, multi-step reasoning that enterprise customers actually need for demanding work.
The context behind this acceleration is an April 2026 renegotiation of the Microsoft-OpenAI deal. Under the new terms, OpenAI can now sell its products through any cloud provider, not just Microsoft's Azure platform. In exchange, Microsoft no longer has to pay OpenAI a revenue share, and crucially, Microsoft gained the formal freedom to build its own models that compete directly with OpenAI. Both companies keep their existing ties: Microsoft remains a 27 percent shareholder in OpenAI and retains a license to OpenAI's technology through 2032. But the deal removed the implicit arrangement where Microsoft had to rely on OpenAI as its sole AI supplier.
The team building these models is led by Mustafa Suleyman, who previously co-founded DeepMind (now part of Google) and AI startup Inflection. Microsoft hired him in 2024 specifically to build this capability. His team has been assembling researchers from DeepMind, academic institutions, and other AI labs.
Alongside MAI-Thinking-1, Microsoft announced six other new models, including an updated image generator now live in PowerPoint, an improved transcription model claimed to outperform Google's and OpenAI's equivalents, and a coding model called MAI-Code-1-Flash that is now available inside the popular developer tool VS Code. Microsoft is also working with the Mayo Clinic on a healthcare-specific model.
For business operators, the practical near-term impact is limited: MAI-Thinking-1 is aimed at enterprise customers using Microsoft's Azure cloud platform, not a consumer feature you'll find in Copilot this week. But the direction is clear. Microsoft is building the infrastructure to power its own products with its own AI, which means less dependency on one external supplier and more room to price, customize, and control what AI does inside Microsoft tools. That eventually flows through to the businesses that use those tools every day.
Microsoft is spending roughly $37.5 billion per quarter on AI infrastructure. The pressure to show that spending produces real, usable products is significant. These model launches are part of that answer.