Microsoft finally has a self-developed speech-to-speech model, and it will set the price first when it launches.

At the Ignite 2026 conference on November 17, Microsoft released MAI Realtime 1. It focuses on real call scenarios in the enterprise: the other party can pick up the call when it is interrupted, confirm when it is time to confirm quickly, and wait until the other party has finished speaking; the environment is noisy, and several people are talking at the same time, and you can still hear what the request is.

When it encounters a task that requires deep thinking, it can call a stronger inference engine and then connect to the enterprise system to complete the task, such as making reservations and handling customer service issues. Tone and style can be adjusted using prompt words.

The price is the toughest part. The basic voice-to-speech fee is US$0.01 per audio minute; if inference is used, it is calculated at US$1 per million input tokens and US$5 per million output.

Comparing data is more intuitive. On the Big Bench Audio benchmark, MAI Realtime 1 has an accuracy of 98.9% and an input audio cost of approximately US$1.35 per hour; GPT-Realtime-2 High is 96.6% and US$4.14 per hour, and Grok Voice 2.0 is 97.2% and costs US$4.80 per hour.

It’s almost one-third the price of its competitors.

It is compatible with the OpenAI Realtime API. Existing voice applications can be migrated without major changes. Agoda and Reservations.ai are already trialling it. Microsoft also plans to put it into the Dynamics 365 customer service center and let it play an AI simulated patient in Dragon Copilot to help nurses practice clinical communication.

At the same conference, Microsoft also released the programming model MAI Code 2. According to Microsoft, its performance is close to similar models from competitors, but its running costs are much lower.

There is also a model specifically geared towards cybersecurity, which will only be available to a small group of customers at first. AI is increasingly being used to find software vulnerabilities, and the pressure on security teams to disclose and respond to vulnerabilities is also rising. Microsoft wants to use this model to help customers catch it.