Microsoft MAI-Thinking-1 Review

MAI-Thinking-1 is the reasoning flagship of Microsoft's first fully in house model family, unveiled at Build 2026 on June 2 and opened to public preview on Microsoft Foundry in August. It is a sparse mixture of experts with roughly one trillion total and 35 billion active parameters and a 256,000 token context window, and unlike Microsoft's earlier reliance on OpenAI it was trained independently on licensed data. It produces an internal chain of thought before answering and scales that effort to the problem.

The significance is strategic as much as technical: Microsoft now has a frontier class reasoner it owns end to end, priced at two dollars in and eight out per million tokens. Self reported numbers are respectable, with 97 percent on AIME 2025 and 52.8 percent on SWE-bench Pro, roughly matching Claude Opus 4.6. What it is not, yet, is a proven leader. It is preview software, the weights are closed, coding trails the top models, and reasoning models are slow and costly for anything high volume.

MAI-Thinking-1 matters most to organisations inside the Microsoft and Azure world who want a first party reasoning model without an OpenAI dependency. For everyone else it is one to watch rather than adopt today: promising, independently trained, and still finding its level against the established frontier.