Microsoft unveils its first advanced reasoning AI model: MAI-Thinking-1
Microsoft introduced its first advanced reasoning AI model, MAI-Thinking-1, at the Build 2026 event. According to The Verge, the new model represents the next level of the flagship Microsoft AI group has developed to date and constitutes a significant step in the company's strategy to reduce dependency on OpenAI.
Microsoft AI group head Mustafa Suleyman, in his introductory presentation on stage, said MAI-Thinking-1 was a 'medium-sized' model, but that it matched leading models on key software engineering benchmarks. Suleyman emphasised that the model was 'trained from the ground up on clean data, without distillation from third-party models.' This statement represents an answer to criticisms that data was carried over from OpenAI in training ChatGPT.
The growth of the Microsoft AI unit over the past two years also reflects the renewed partnership terms with OpenAI. At the end of last year, the two companies signed a new agreement loosening the business partnership previously announced to run until 2030. This agreement allows Microsoft to develop its own models and OpenAI to make its own independent infrastructure investments.
The technical specifications of MAI-Thinking-1 have not yet been fully shared with the public. Microsoft AI chief product officer Erik Meijer, speaking to The Verge, said 'although the model has a low parameter count, it has been trained in such a way as to compete with leading models on reasoning tasks.' Meijer emphasised that MAI-Thinking-1 was particularly strong on software engineering tasks, code generation and debugging.
Other models introduced at Build 2026 include MAI-Voice-1 (voice AI), MAI-Coder-1 (a small model focused on code writing) and MAI-Vision-1 (a visual understanding model). The model family presented by Microsoft AI will be made available directly to customers on the company's Azure cloud platform. Microsoft chief executive Satya Nadella, in his presentation at the event, said 'diversity is now needed for our customers to decide which model to use.'
One of the sectoral impacts of the model could be Microsoft's entering into competition with OpenAI on the strategy of productising its own models. Speaking to The Verge, independent AI expert MIT Professor Hartmut Neven said 'Microsoft's own model is an important step in the name of owning AI infrastructure; but in terms of data centre and training costs, an additional 8-10 billion dollars of investment per year is needed to be able to compete with OpenAI.'
Microsoft Copilot products will offer a new option selection with MAI-Thinking-1. Until now, Copilot Pro subscriptions have used OpenAI's GPT models by default. MAI-Thinking-1 is planned to be integrated into the Copilot product range in early 2027. Microsoft chief revenue officer Judson Althoff said 'we will provide intelligent routing between different models for users.'
In terms of specific benchmark results, MAI-Thinking-1 showed a 67 percent success rate on SWE-Bench software engineering tasks, 84 percent accuracy on the MMLU general knowledge test, and 92 percent accuracy on the GSM8K mathematics test. These figures are at a level close to Anthropic's Claude Opus 4.7 model; but slightly behind OpenAI's o3 model's 71 percent SWE-Bench score.
On the Microsoft Azure platform, MAI-Thinking-1's consumption model is at a price point 22 percent lower than existing OpenAI products. The company forecasts the potential revenue stream coming from Azure at 4.5 billion dollars for the 2026 financial year. Microsoft chief financial officer Amy Hood, on the last investor call, said 'revenue from AI products will provide significant contribution to the growth rate of the cloud segment.'
This new phase of Microsoft's AI strategy has also drawn comment in the global tech investment market. Wedbush analyst Dan Ives said 'Microsoft's own model is a building block that will turn the Copilot product line into a lasting category.' Forrester researcher J.P. Gownder said enterprise customers would view MAI-Thinking-1 as 'a critical point of validation' for preferring Microsoft's end-to-end AI solution. This article does not constitute investment or product-purchase advice.