MAI Realtime 1

MAI Realtime 1 is Microsoft AI’s first speech-to-speech model, announced at Microsoft Ignite on 2026-11-17 for enterprise voice applications. It takes audio in and returns audio out. Part of the MAI series.

Capabilities (Microsoft’s post)

  • Conversation: handles interruptions, quick acknowledgements and turn-taking.
  • Adaptive reasoning: a fast path that escalates to a stronger reasoning engine for harder tasks.
  • Noise and speakers: recognises requests with background noise or several speakers.
  • Actions: connects to enterprise systems and tools.
  • Personas: tone and behaviour set through prompting.
  • Compatibility: works with the OpenAI Realtime API, which Microsoft says eases testing and migration.

Supported languages are not stated.

Availability

  • MAI Labs: a trial at the Labs site.
  • Microsoft Foundry: limited-access preview by registration. Microsoft says preview customers will not be billed until public preview begins.
  • General availability: no date is given.
  • Named customers: Agoda and Reservations.ai.
  • Planned integrations: Dynamics 365 Contact Center and Dragon Copilot.

Billing

Base speech-to-speech is billed per audio minute. Reasoning is billed separately per token. The post gives the amounts; they are not recorded here, so check the post or Microsoft’s pricing page.

Microsoft’s claims

  • Outperforms GPT-Realtime 2.1 “on quality, speed, and price across public benchmarks and Microsoft product evaluations”.
  • Sits on the performance-price Pareto frontier for Big Bench Audio and Tau-Voice.
  • No latency figures or benchmark scores are given in the text reviewed.

Source limits

  • The claims are Microsoft’s, measured under configurations the post does not fully describe.
  • Languages and GA timing are not stated.

Sources