MAI Realtime 1 is Microsoft AI's first speech-to-speech model, announced on November 17, 2026, at Ignite. It is built for voice agents that hold natural conversations and complete tasks for customers. The base price starts at $0.01 per audio minute.
Microsoft says the model beats OpenAI's GPT-Realtime 2.1 on quality, speed and price. The model is in preview for now, and those comparisons come from Microsoft itself.
What MAI Realtime 1 does
A speech-to-speech model listens to audio and answers with audio directly, without a separate transcription and voice step in between. That makes it suited to phone lines, contact centers and voice assistants, where delays and awkward turn-taking break the conversation.
Microsoft lists the main features. The model handles interruptions and short acknowledgments like "mm-hm," and it waits for the speaker to finish. It is designed to understand requests despite background noise or other people talking. It can also call tools and connect to enterprise systems, so a voice agent can act as well as answer.
The model also works in two speeds. A fast system keeps the conversation flowing in real time. When a request needs deeper thinking, it hands the work to a stronger reasoning engine without long pauses. Developers can change the agent's tone, style and persona through prompts.
For developers, the most practical detail is compatibility. MAI Realtime 1 works with the OpenAI Realtime API, so teams can test it in existing voice apps without rewriting them. That makes the comparison with OpenAI's own model the obvious next question.
MAI Realtime 1 vs GPT-Realtime 2.1
Microsoft says MAI Realtime 1 outperforms GPT-Realtime 2.1 on quality, speed and price. It tested both models on two public benchmarks. Big Bench Audio checks spoken reasoning, and τ-Voice measures how well a voice agent completes multi-step tasks. Microsoft also used internal tests from Dynamics 365 Contact Center.
The model page lists two figures. MAI Realtime 1 answers 98.9% of the spoken reasoning questions in Big Bench Audio correctly. Its median time to the first audio response on that test is 880 milliseconds.
Microsoft has not published GPT-Realtime 2.1's scores next to these, or a τ-Voice score for its own model. It says only that MAI Realtime 1 sits on the performance-price frontier for both benchmarks. Artificial Analysis, which runs an independent speech-to-speech leaderboard, has not listed the model yet.
The price gap is easier to check.
MAI Realtime 1 pricing
The base speech-to-speech service costs $0.01 per minute, or 60 cents per hour of conversation. Reasoning is billed on top, at $1 per million input tokens and $5 per million output tokens. The final cost per call therefore depends on how often the agent needs to think hard.
OpenAI prices GPT-Realtime 2.1 by tokens instead of minutes. Audio costs $32 per million input tokens and $64 per million output tokens, and text costs $4 and $24. The two pricing models do not convert one to one. Still, a flat cent per minute is aggressive for enterprise voice, where calls run for millions of minutes a month.
Preview customers will not pay anything until the public preview begins, and the model is already reaching real products.
Where MAI Realtime 1 is available
Developers can try MAI Realtime 1 in MAI Labs, Microsoft AI's demo site, or register for a limited-access preview in Microsoft Foundry. Travel platform Agoda and Reservations.ai are already testing it.
Microsoft is also building it into its own products. It is coming to Dynamics 365 Contact Center for customer service. In Dragon Copilot, Microsoft's clinical assistant, it will power simulated patient conversations that help nurses practice.
The voice model is one of several models Microsoft AI launched at Ignite. The lineup also includes the MAI Code 2 coding model in GitHub Copilot and the MAI Image 2.7 image editor. All of them follow the same pitch: near-frontier quality at a much lower price.
Is MAI Realtime 1 better than GPT-Realtime 2.1
Microsoft says yes, but the evidence is thin so far. The company has published one benchmark score and one latency figure for its own model. It has shown no side-by-side numbers for OpenAI's model, and no independent leaderboard has tested it.
The price is the stronger argument. At $0.01 a minute plus reasoning, MAI Realtime 1 is built to undercut voice agents on token-based pricing. Because it speaks the OpenAI Realtime API, switching costs little to try. Teams running large contact centers have a good reason to test it once the public preview opens.
FAQ
Here are short answers to common questions about the launch.
What is MAI Realtime 1?
MAI Realtime 1 is Microsoft AI's first speech-to-speech model. It powers voice agents that talk with users in real time, handle interruptions and call tools.
How much does MAI Realtime 1 cost?
The base price is $0.01 per audio minute. Reasoning costs extra, at $1 per million input tokens and $5 per million output tokens.
Can I use MAI Realtime 1 now?
You can try it in MAI Labs or register for the limited-access preview in Microsoft Foundry. Preview customers are not billed until the public preview starts.
Sources
- MAI Realtime 1: Leading voice AI at $0.01 per audio minute, Microsoft AI – November 17, 2026
- MAI Realtime 1 model page, Microsoft AI
- Seven new world class models at scalable price-performance, Microsoft AI – November 17, 2026
- GPT-Realtime-2.1 model page, OpenAI API
- Microsoft AI on X – November 17, 2026