Microsoft's MAI Realtime 1 Is Its First Live Speech-to-Speech Model, and It Bills Reasoning Apart From Audio

Microsoft AI introduced MAI Realtime 1 at Ignite on November 17, 2026. The speech-to-speech model is compatible with the OpenAI Realtime API, starts at $0.01 per audio minute with reasoning billed separately at $1/$5 per million tokens, and is heading to Dynamics 365 Contact Center and Dragon Copilot. Quality comparisons with GPT-Realtime 2.1 are Microsoft's own.

A ship captain speaks into a brass speaking tube while a navigator in the cabin below answers through it, a small hourglass running on the chart table beside him.
A ship captain speaks into a brass speaking tube while a navigator in the cabin below answers through it, a small hourglass running on the chart table beside him.

Microsoft now has its own model for live voice conversations. At Ignite on November 17, 2026, Microsoft AI introduced MAI Realtime 1, which it describes as its first speech-to-speech model, built for natural, low-latency conversations in enterprise applications. It was one of seven models Microsoft AI released that day, alongside new coding, cyber-defence, image and decision models.

Until now, Microsoft’s in-house audio work has come in pieces: MAI Voice 2 Flash for speech synthesis and MAI Transcribe 2 for speech recognition. MAI Realtime 1 is meant to handle the live conversation itself, the part where a voice agent has to listen, reason, talk and act at the same time.

A fast talker backed by a slower thinker

Microsoft lists five capabilities. The model is meant to handle interruptions, follow quick acknowledgements and wait for a speaker to finish. It is designed to recognise requests through background noise or other people talking. It connects to enterprise systems and tools so an agent can act rather than only answer. And it adjusts tone, style and behaviour through prompting to suit a domain or customer.

Microsoft also describes what it calls adaptive intelligence. A fast, responsive system keeps the conversation going in real time, and when a task needs deeper thought it “taps into a more powerful reasoning engine” so the caller does not hear awkward pauses. Other voice stacks use a similar division of labour; Baseten’s reference voice agent, for example, pairs a fast co-located “talker” model with a stronger “thinker” that handles reasoning and tool calls in the background.

The split shows up in the pricing. Base speech-to-speech usage costs $0.01 per minute. Reasoning is billed separately at $1 per million input tokens and $5 per million output tokens. A simple call that never needs the reasoning engine stays close to the per-minute rate, while calls that lean on it cost more.

Built to drop into existing voice apps

MAI Realtime 1 is compatible with the OpenAI Realtime API, which Microsoft says makes it straightforward to test the model and migrate existing voice applications.

Microsoft says MAI Realtime 1 beats GPT-Realtime 2.1 on quality, speed and price across public benchmarks and Microsoft product evaluations. It names Big Bench Audio, which measures spoken reasoning, and τ-Voice, which measures agentic task completion over voice, and says the model sits on the price-performance frontier for both. It also cites Dynamics 365 Contact Center evaluations. The announcement does not publish the underlying scores, and Microsoft notes that results reflect specific evaluation configurations and that latency and cost comparisons vary by application.

A day earlier, Baseten opened a private preview of Baseten Realtime Voice, which also speaks the OpenAI Realtime protocol and is priced at about a cent per audio minute at scale.

Contact centres and clinical training first

Microsoft says Agoda and Reservations.ai are already testing the model. The company is bringing MAI Realtime 1 to Dynamics 365 Contact Center for customer service, and to Dragon Copilot, where it will run AI-simulated patient conversations that help nurses practise clinical interactions.

Those destinations match Microsoft’s earlier audio rollouts. MAI Voice 2 Flash already powers voices in Dynamics 365 Contact Center, which Microsoft said in July cut GPU costs by up to 89%. At Ignite, the company also said Dragon Copilot Transcribe, built on MAI-Transcribe-2, now handles 100% of doctor-patient conversation transcription in Dragon Copilot. Microsoft says that covers 10 million encounters a month, halves error rates in multilingual workflows and reaches a third of US physicians.

MAI Labs for individuals, a Foundry preview for businesses

MAI Realtime 1 is not yet generally available. Individuals can try it in MAI Labs, Microsoft’s limited preview for experimental models, which is restricted to personal, non-commercial use. Businesses can register for a limited-access preview in Microsoft Foundry. Microsoft says preview customers will not be billed until public preview begins, and it has not given a date for that or for general availability.

Microsoft has not said which languages MAI Realtime 1 supports, how large it is or what latency it reaches.

Topics
  • #Microsoft
  • #MAI Realtime 1
  • #Speech-to-Speech
  • #Voice Agents
  • #OpenAI Realtime API
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.