Microsoft developing AI reasoning models to compete with OpenAI

Microsoft is working on a new generation of in-house artificial intelligence reasoning models designed to rival those from OpenAI, according to a report by The Information.

The company, which has been a major backer of OpenAI, is exploring options to reduce its dependence on external AI technology and may eventually offer these models to developers via an application programming interface.

The report indicates that Microsoft is testing models from xAI, Meta, and DeepSeek as potential replacements for the technology currently powering its Microsoft 365 Copilot. Reuters reported in December that Microsoft had been actively integrating both internal and third-party AI models into Copilot to diversify its technological foundation and lower costs, given that the service was originally built on OpenAI's GPT-4.

Microsoft’s AI division, led by Mustafa Suleyman, has reportedly completed training a family of models internally known as MAI. These models are said to perform nearly as well as leading models from OpenAI and Anthropic on standard benchmarks. In addition, Suleyman’s team is developing reasoning models that use chain-of-thought techniques—methods that generate answers through intermediate reasoning steps—to directly compete with similar models from OpenAI.

Suleyman’s team is already experimenting with replacing OpenAI’s models in Copilot with the newly developed MAI models, which are significantly larger than an earlier series known as Phi. Microsoft is considering a release of these models later this year in the form of an API, which would enable external developers to integrate the advanced AI reasoning capabilities into their own applications.

Neither Microsoft nor OpenAI immediately responded to requests for comment.

Copied