Overview
MAI Realtime 1 is Microsoft AI's first speech-to-speech model, announced at Ignite on November 17, 2026. It is built for flowing, real-time conversations in enterprise voice agents: it handles interruptions, follows quick acknowledgements, waits for a speaker to finish, copes with background noise and other voices, and connects to enterprise systems and tools so an agent can act on a request rather than only answer it. Tone, style and behaviour are set through prompting.
A fast responsive system keeps the conversation going while a separate, more powerful reasoning engine is called for harder requests. Microsoft prices base speech-to-speech use at $0.01 per minute and bills reasoning separately at $1 per million input tokens and $5 per million output tokens. The model is compatible with the OpenAI Realtime API, so existing voice applications can be pointed at it.
Microsoft says it outperforms GPT Realtime 2.1 on quality, speed and price across public benchmarks and Microsoft product evaluations, and sits on the performance-price Pareto frontier for Big Bench Audio speech reasoning and τ-Voice agentic task completion. In Microsoft's internal evaluation it scored 98.9% on a 40-question Big Bench Audio subset and 57.5% τ-Voice task completion, each at $1.3548 per hour of input audio; competitor figures in Microsoft's charts come from Artificial Analysis.
At launch it can be tried in MAI Labs or through a limited-access preview in Microsoft Foundry, and preview customers are not billed until public preview begins. Microsoft names Agoda and Reservations.ai as early testers and is bringing the model to Dynamics 365 Contact Center and to Dragon Copilot, where it supports AI-simulated patient conversations for nurses' clinical practice.
| Released | 2026-11-17 |
|---|---|
| License | Proprietary |
| Weights | API only |
| Modalities | Audio |
| Status | Preview |
Benchmarks
Scores on a 0–100 scale (25-point gridlines); higher is better. Each benchmark links to its published source.
Pricing
| Input | $0.01 / audio minute |
|---|
Base speech-to-speech price; reasoning is billed separately at $1 per 1M input tokens and $5 per 1M output tokens. Preview customers are not billed until public preview.
Strengths
- Base speech-to-speech price of $0.01 per audio minute, with reasoning billed separately at $1 / $5 per million input / output tokens
- Compatible with the OpenAI Realtime API, easing migration of existing voice apps
- 98.9% on a Big Bench Audio subset and 57.5% τ-Voice task completion in Microsoft's internal evaluation
- Handles interruptions, quick acknowledgements and turn-taking, and understands speech over background noise
- Hands harder requests to a more powerful reasoning engine without pausing the conversation
Best for
- Reach for it for customer-service voice agents, the role Microsoft is bringing it to in Dynamics 365 Contact Center.
- Reach for it to move an app built on the OpenAI Realtime API to a lower per-minute price.
- Reach for it for simulated conversations, such as the practice patient conversations for nurses in Dragon Copilot.
- Look elsewhere for production traffic today if you need general availability — at launch it is in MAI Labs and a limited-access Foundry preview.
How to access
| Provider | Model ID |
|---|---|
| MAI Labs ↗ | — |
| Microsoft Foundry (limited-access preview) ↗ | — |
FAQ
What is MAI Realtime 1?
It is Microsoft AI's first speech-to-speech model, announced on November 17, 2026, for real-time voice agents that hold natural conversations and act on enterprise systems.
How much does MAI Realtime 1 cost?
Base speech-to-speech use is $0.01 per minute; reasoning is billed separately at $1 per million input tokens and $5 per million output tokens. Preview customers are not billed until public preview begins.
Does MAI Realtime 1 work with the OpenAI Realtime API?
Yes. Microsoft says it is compatible with the OpenAI Realtime API, so developers can test it and migrate existing voice applications.
How can I try MAI Realtime 1?
Try it in MAI Labs, or register for the limited-access preview in Microsoft Foundry.
