Microsoft releases 7 MAI models, with voice from $0.01 a minute
Microsoft tells that its MAI Code 2 model gives 98% of the performance of GPT-6.1 Sol at 40% of the price, in its own tests.
At Ignite, Microsoft released a family of 7 MAI models. They are for live voice, code, cyber defense, images, voice transcription and model routing. Microsoft tells that they give good performance at a lower price than models of other companies. Customers can also operate a MAI model with their own data in their own Azure cloud, as MAI Licensed Weights.
What is new
These models are in the family:
- MAI Realtime 1: a live voice model, from $0.01 for each audio minute.
- MAI Code 2 and Code 2 Flash: tasks of code each day, in GitHub Copilot.
- MAI Cyber 2 Flash: security scans, with output tokens 86% cheaper than the earlier Flash model.
- MAI Image 2.7: image editing, ranked first on Artificial Analysis.
The proof
All tests are from Microsoft. In a blind test of front end work, developers preferred MAI Code 2 to GPT-6 Sol 68% of the time. Microsoft tells that results change with the workload and the method.
What is not known
Microsoft tells that the models are not tuned for hard science or for very long agent work. It gave no external test of the code claims.
Insight: Microsoft is the first customer of its own cheap models
Microsoft tells that MAI Cyber is almost 40% of the scans in its own security system. On 12 November 2026 it released a Flash model that costs 18% less to operate than version 1.0. The next one is 86% cheaper on output tokens. This seems to be a plan to prove cost cuts at home, then sell the cheap models to others. Test the price on your own tasks.
Sources
Posted