Back home

MAI Realtime 1

sitemap+web 0.93
Visit source

Intro

MAI Realtime 1 is Microsoft AI’s first speech-to-speech model, introduced on 17 November 2026 for naturally flowing enterprise voice conversations, starting at $0.01 per audio minute. It belongs to the same Ignite model wave as MAI Code 2, MAI Cyber 2 Flash, and the already filed MAI Image 2.7, but it is a new real-time voice line rather than an alias of a text or image SKU.

Microsoft compares it with GPT Realtime 2.1 and says it outperforms that model on quality, speed, and price across public benchmarks and Microsoft product evaluations, sitting on the performance-price Pareto frontier for Big Bench Audio (speech reasoning) and τ-Voice (agentic task completion). A pricing footnote states that base speech-to-speech usage is $0.01 per minute, with reasoning billed separately at $1 per million input tokens and $5 per million output tokens.

The official getting-started path is to try it in MAI Labs or register for a Foundry limited-access preview; preview customers are not billed until public preview begins. The post also says the model is compatible with the OpenAI Realtime API, so existing voice apps can be tested and migrated more easily.

Advertisement

Explain

The capability list is about conversation that can also take action: handling interruptions and brief acknowledgments, and knowing when to wait; keeping the exchange fluid with a low-latency system, then calling a stronger reasoning engine when a turn needs deeper thought, without an awkward pause; recognizing requests through background noise or other speakers; connecting to enterprise systems and tools so a voice agent can do more than answer questions; and adapting tone, style, and behavior through prompting for different customers, domains, and use cases.

On applications, Microsoft names Agoda and Reservations.ai as early testers, and says it is bringing MAI Realtime 1 to Dynamics 365 Contact Center for customer service and to Dragon Copilot for AI-simulated patient conversations that help nurses practice clinical interactions. Those are launch-time partnerships and product plans, separate from the benchmark claims. The footnotes also warn that performance, latency, and cost vary by application and methodology.

Usage

According to the official getting-started note, try MAI Realtime 1 in MAI Labs or register for the Microsoft Foundry limited-access preview; preview usage is not billed until public preview begins. If you already have a voice app on the OpenAI Realtime API, Microsoft says this model is compatible with that interface for evaluation and migration.

Billing should be read as a base per-minute speech charge plus separately billed reasoning tokens. Dynamics 365 Contact Center and Dragon Copilot are Microsoft’s own product paths; their timing and eligibility follow those products’ docs. Do not confuse this with MAI Transcribe 2, which already existed in September as a speech-to-text model, or with the separately filed MAI image and coding SKUs: Realtime 1 is explicitly described as Microsoft’s first speech-to-speech model.