Microsoft launches MAI Realtime 1 voice AI at one cent per minute

MAI Realtime 1 is Microsoft's first speech-to-speech model. It's built to hold conversations, handle interruptions and get work done in business software while it talks. Base usage starts at $0.01 per audio minute. It's one of seven new models Microsoft just released.

Model releases Nov 18, 2026 7 min read Source: Bing News (en-US)
Illustrative image, AI-generated

What's new

Microsoft has unveiled MAI Realtime 1, its first in-house speech-to-speech model. That's an AI that listens to spoken language and answers out loud. It's designed for natural real-time conversations, such as voice assistants in customer service, automated phone switchboards or digital receptionists. Pricing starts at $0.01 per audio minute.

According to Lapaas Voice, MAI Realtime 1 launched alongside six other models covering coding, cybersecurity, image generation, transcription and automatic model selection. For now, the model is only available through Microsoft's MAI Labs and a limited preview in Microsoft Foundry, Microsoft's platform for AI developers. Microsoft says preview customers won't be charged until the public preview begins.

What MAI Realtime 1 is supposed to do

Traditional voice assistants usually work in three separate steps. First, speech gets converted to text. Then a language model processes that text. Finally, a synthetic voice reads the answer aloud. Every step adds delay, and the conversation ends up feeling choppy, like talking over a walkie-talkie. MAI Realtime 1 is meant to merge these steps into one continuous flow. Microsoft lists five core capabilities:

  • Natural conversation: The model copes with interruptions and quick interjections like "mm-hmm" and is meant to recognize when someone has finished speaking. You don't have to wait for rigid turn-taking.
  • Adaptive intelligence: It answers simple requests quickly. For complex tasks it switches to deeper reasoning, thinking step by step before it answers.
  • Audio understanding: It's designed to understand speech even with background noise or several people talking.
  • Real actions: Connected to company systems, the model is supposed to complete tasks rather than just answer questions, for example booking an appointment.
  • Customizable behavior: Developers can set the voice's tone, style and personality.

Pricing: cheap at the base rate, reasoning costs extra

The price is the headline. Plain speech processing is cheap, but reasoning is billed separately by tokens, the small chunks of text that language models work and bill in.

Price component Cost
Speech-to-speech base usage $0.01 per audio minute
Reasoning input $1 per million tokens
Reasoning output $5 per million tokens

The report gives an example: a voice system handling 10,000 minutes of conversation costs about $100 in base usage, with reasoning on top. So what a deployment really costs depends heavily on how often, and how much, the model has to think.

The OpenAI comparison and easy switching

Microsoft claims MAI Realtime 1 beats OpenAI's GPT Realtime 2.1 on quality, speed and price. It bases this on public benchmarks, which are standardized test suites. Big Bench Audio tests reasoning on spoken tasks, and Tau-Voice measures how well a voice agent handles multi-step tasks. Microsoft also ran internal tests in its Dynamics 365 Contact Center customer service product. It says its model sits on the "price-performance frontier", meaning no better model exists at that price. Nobody has tested this independently yet, and the report itself points out that the comparisons rely on Microsoft's own test setups.

The strategically important part is compatibility with the OpenAI Realtime API. An API is the programming interface apps use to talk to a model. If you've already built a voice app on OpenAI's interface, you can try Microsoft's model without rebuilding everything. That makes it easier for companies to switch providers.

Early customers and Microsoft products

Microsoft is pitching the model mainly at businesses. The company says travel platform Agoda and Reservations.ai are already testing it. MAI Realtime 1 is also coming to Microsoft's own products:

  • Dynamics 365 Contact Center: Voice agents answer customer questions, give out information and take over routine service calls.
  • Dragon Copilot: In Microsoft's healthcare product, the model plays the patient so nurses can practice clinical conversations.

The six other models

MAI Realtime 1 is just one part of the announcement. Instead of one general-purpose model for everything, Microsoft is betting on specialized models for specific jobs:

Model Purpose
MAI Code 2 Demanding coding and software engineering tasks
MAI Code 2 Flash Faster, cheaper coding tasks
MAI Cyber 2 Flash Cybersecurity and threat analysis
MAI Image 2.7 Image generation and editing
MAI Realtime 1 Real-time voice conversations
MAI Transcribe 2 Multilingual transcription
Decision 1.1 Routing: deciding which model handles a request

The main source gives no benchmarks or prices for the other six models.

Background: Microsoft's voice toolkit

The new model didn't come out of nowhere. Microsoft started building its own models in 2025 with MAI-Voice-1 and MAI-1 Preview, aiming to rely less on OpenAI's GPT models. In fall 2026 it released several audio models, as reported by unite.ai and Tech Times:

  • On September 3, Microsoft released MAI-Transcribe-2, a batch transcription model. It processes finished audio files and costs $0.10 per audio hour at its introductory price.
  • On October 1, MAI-Transcribe-2-Streaming followed. It transcribes speech while someone is still talking, in 60 languages and with automatic language detection. Microsoft says it delivers its first provisional results, known as partials, after just over 100 milliseconds. The word error rate is said to be 2.5 percent. The introductory price is $0.54 per hour.
  • At the same time, Microsoft released the text-to-speech models MAI-Voice-2.1 at $22 per million characters and MAI-Voice-2.1-Flash at $15 per million characters. They support 23 languages, voice cloning from a few seconds of reference audio, and a voice that sounds the same across languages.

Tech Times calculated back then that a voice agent built from these pieces could complete a conversational turn in well under a second: about 130 milliseconds to listen, around 350 to think and 150 until speech starts. But the magazine was skeptical. When it wrote the article, Microsoft's top spot in Artificial Analysis's streaming ranking hadn't been independently confirmed, and Grok Voice Transcribe 2.0 was still in first place on the public leaderboard. Tech Times also noted that benchmark word error rates often get much worse in real-world use. On top of that, all of these models were labeled public preview with no service level agreement, meaning no contractually guaranteed availability. MAI Realtime 1 looks like the next step: one model handles the whole pipeline instead of three building blocks chained together.

Our take

With MAI Realtime 1, Microsoft is going head-to-head with OpenAI's realtime models, even though OpenAI was long its closest AI partner. The compatible API makes the intent clear: Microsoft wants to make switching as easy as possible for OpenAI customers. A base price of one cent per minute is attractive for call centers handling thousands of calls a day. But the total cost depends on reasoning, which is billed separately, so a direct cost comparison with OpenAI is tricky.

A lot is still unclear. Every performance comparison comes from Microsoft itself, and there are no independent measurements of quality, latency or error rates. It's also unclear which languages MAI Realtime 1 supports and how well it handles German, since the source doesn't say. Individual users can barely get at the model for now, because access is limited to MAI Labs and a restricted developer preview. The weights aren't openly available, and the model runs only in Microsoft's cloud. It matters most to companies that want to run voice agents at scale, and as a sign that Microsoft is increasingly going its own way in the AI race.

Sources

#microsoft#voice ai#voice-agents#openai#customer service
Read the original source

Ask the AI assistant

Questions about this article, installation or background? Our assistant knows this page and runs on our own machine, not in someone else's cloud.

Connecting …

The Assistant is an AI (Gemma 4) and can make mistakes – please check important details against the source. Questions are used only to generate an answer and are deleted after one hour. Please don't enter any personal data.

Comments

What do you think? Already tried it? Write it here, no sign-up needed.

  1. Loading comments …

Comments are public. Links aren't allowed, and we delete abusive posts. You can delete your own comment at any time in this browser. Learn more in the Privacy Policy.