Microsoft releases MAI models to lower AI cost

Microsoft starts to release foundation models that it makes, with the name MAI, in its Foundry platform. A foundation model is a large AI model that does many tasks, and Foundry also has more than 11,000 models from other companies. The MAI models are made to give a lower token cost than competitors have, and a token is a small piece of text.

How it works

MAI-Thinking-1 is the model for the most complex tasks. Teams can use it through the Foundry Azure OpenAI APIs (software connections) that they have. They only change the endpoint address in their code. A team can set a token limit for each task.

The numbers

MAI-Cyber-1.1-Flash does 90% of the tasks in the MDASH security system. It sends a task that it cannot do to a larger model. Microsoft shows benchmark results. The source tells that some results are higher for large models from Anthropic and OpenAI, but it gives no figures.

What is not known

The source gives no price for each token. MAI-Code-1.1-Flash must have more than 120GB of RAM (computer memory), and does not operate on Copilot+ PCs. Teams must compare a device with a high cost to the cloud for their agent use.

Insight: Microsoft makes its model easy to swap in

Foundry has more than 11,000 models, and the MAI models use the same APIs as Azure OpenAI. In October 2026 Microsoft also added MAI-Code-1.1-Flash to Foundry tuning. This reads as a plan to be the first model that Foundry teams use. Monitor if the agent router selects MAI models first.