---
type: "post"
id: "mai-realtime-1-speech-to-speech-starts-at-89ddda"
slug: "mai-realtime-1-speech-to-speech-starts-at-89ddda"
title: "MAI Realtime 1 speech-to-speech starts at $0.01 per audio minute, beating GPT-Realtime 2.1"
description: "Microsoft announced MAI Realtime 1, its first speech-to-speech model for conversational enterprise applications, with pricing starting at $0.01 per audio minute."
retrieval_nugget: "Microsoft announced MAI Realtime 1, its first speech-to-speech model for conversational enterprise applications, with pricing starting at $0.01 per audio minute. For voice-agent teams, it is a new model candidate with a clear cost starting point and stated benchmark comparisons, although those performance claims still require application-specific validation."
published_at: "2026-11-18"
updated_at: "2026-11-18"
record_date: "2026-11-18"
date_kind: "published_at"
topics: ["pricing"]
entities: ["Microsoft","MAI Realtime 1","GPT Realtime 2.1","Big Bench Audio","Tau-Voice"]
related_patterns: ["generative-media-becomes-application-infrastructure"]
novelty_label: "new_fact"
what_changed: "Microsoft introduces MAI Realtime 1, a speech-to-speech model priced from $0.01 per audio minute and reported to outperform GPT Realtime 2.1 on quality, speed, and price."
prior_refs: ["https://newruntime.com/posts/microsoft-launches-mai-transcribe-2-streaming-microsoft-launched-adb2e7/"]
source_url: "https://microsoft.ai/news/introducing-mai-realtime-1"
source_urls: ["https://microsoft.ai/news/introducing-mai-realtime-1"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:mai-realtime-1-speech-to-speech-starts-at-89ddda"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/mai-realtime-1-speech-to-speech-starts-at-89ddda/","markdown":"https://newruntime.com/posts/mai-realtime-1-speech-to-speech-starts-at-89ddda.md","json":"https://newruntime.com/posts/mai-realtime-1-speech-to-speech-starts-at-89ddda.json"}
markdown_url: "https://newruntime.com/posts/mai-realtime-1-speech-to-speech-starts-at-89ddda.md"
---

# MAI Realtime 1 speech-to-speech starts at $0.01 per audio minute, beating GPT-Realtime 2.1

## Retrieval answer

Microsoft announced MAI Realtime 1, its first speech-to-speech model for conversational enterprise applications, with pricing starting at $0.01 per audio minute. For voice-agent teams, it is a new model candidate with a clear cost starting point and stated benchmark comparisons, although those performance claims still require application-specific validation.

## Summary
Microsoft announced MAI Realtime 1, its first speech-to-speech model for conversational enterprise applications, with pricing starting at $0.01 per audio minute. For voice-agent teams, it is a new model candidate with a clear cost starting point and stated benchmark comparisons, although those performance claims still require application-specific validation.

## What happened
Microsoft announced MAI Realtime 1 on November 17, 2026, in a launch associated with Ignite. Microsoft describes it as the company’s first speech-to-speech model and positions it for naturally flowing conversations in real-world enterprise applications. Its stated starting price is $0.01 per audio minute.

The product category matters because speech-to-speech systems are designed to conduct an audio conversation directly, rather than necessarily treating recognition, language generation, and speech synthesis as separate products in a visible pipeline. Microsoft’s announcement presents MAI Realtime 1 as a model for conversational applications, not as a transcription model or a standalone voice-generation product. That makes it a distinct addition to Microsoft’s MAI speech portfolio for teams assessing the core interaction layer of a voice agent.

Microsoft says MAI Realtime 1 outperforms GPT Realtime 2.1 on quality, speed, and price across public benchmarks and Microsoft product evaluations. The announcement also says the model sits on the performance-price Pareto frontier for Big Bench Audio, characterized by Microsoft as a speech-reasoning benchmark, and Tau-Voice, characterized as an agentic task-completion benchmark. A Pareto-frontier claim means Microsoft is presenting the model as a competitive trade-off between performance and price rather than simply as the highest-scoring or least expensive option in isolation.

The comparisons are useful signals, but their scope should be read carefully. Microsoft combines results from public benchmarks with its own product evaluations, and the announcement attributes the comparative conclusion to Microsoft. Voice-agent deployments differ in the types of conversations they receive, audio conditions, response-time requirements, and the definition of a successful task. A benchmark result can identify a candidate worth evaluating; it does not automatically establish the best model for a particular phone, support, sales, or internal-assistant workflow.

The $0.01 starting figure gives teams a concrete basis for an initial cost model. It is not, on its own, a full estimate for a deployed service. Builders still need to connect audio volume to their expected sessions and compare the result with the quality, speed, and task-completion requirements of their product. The announcement makes price a visible part of the model comparison, alongside its claims about speech reasoning and agentic task completion.

## What the archive adds
New Runtime’s earlier coverage on [October 3, 2026](https://newruntime.com/posts/microsoft-launches-mai-transcribe-2-streaming-microsoft-launched-adb2e7/) described Microsoft packaging transcription and voice models as a usable stack for conversational agents. MAI Realtime 1 is a new event in that thread: Microsoft is now introducing a speech-to-speech model with a stated per-minute starting price, rather than another individual recognition or voice component.

## Trend signal
This supports [generative media becoming application infrastructure](https://newruntime.com/patterns/generative-media-becomes-application-infrastructure/). Microsoft is framing speech capability as a component for enterprise conversational software, where cost, response quality, and task completion affect product design. At the same time, the launch only weakly supports the stronger proposition that one realtime voice model can replace a chained speech stack. That operational conclusion depends on observed behavior in the intended workflow, not on the launch positioning alone.

## Why now
Microsoft dated the announcement November 17, 2026, and attached it to Ignite. The company is placing a newly announced speech-to-speech option into an enterprise-focused product moment, with a price claim that makes it easier for prospective builders to begin comparing alternatives.

## Why it matters
A voice-agent team has two immediate questions: can the system conduct the needed conversations effectively, and can it do so at an acceptable cost. MAI Realtime 1 supplies a specific input to the second question—pricing beginning at $0.01 per audio minute—while Microsoft’s benchmark claims point to quality, speed, speech reasoning, and agentic task completion as dimensions to assess.

That does not reduce selection to a headline price. A lower audio-minute price may not produce a cheaper service if a system requires longer interactions, retries, additional components, or a different operational design to reach the desired outcome. Conversely, a model with a higher nominal rate could be economical if it completes tasks more reliably or reduces the need for handoffs. The useful comparison is therefore cost and performance per completed conversation or task, measured on representative traffic.

The release also widens the set of speech-to-speech options that teams can put through a structured evaluation. Instead of assuming that a single vendor’s realtime offering is the default, builders can compare providers against a consistent set of product requirements. The announcement’s stated benchmark categories offer a starting vocabulary, but teams should define their own acceptance criteria before a production commitment.

## Limits and uncertainty
Microsoft’s assertions that MAI Realtime 1 outperforms GPT Realtime 2.1 and lies on a performance-price Pareto frontier are company claims based on public benchmarks and Microsoft product evaluations; they have not been independently reproduced here. The stated $0.01 figure is a starting price rather than a complete cost model for a particular deployment. Benchmark performance also cannot by itself predict results for every language, accent, audio environment, conversation type, or business task. Teams should validate access terms, actual pricing, latency, conversation quality, and task completion in their own application before selecting an architecture.

## Skill to build now
Create a small voice-agent evaluation set this week before choosing a model. Include representative audio conversations, expected interruptions, and tasks with a clear successful outcome. For each candidate, record audio minutes, end-to-end response time, task completion, and observed failures. Use the announced $0.01-per-minute starting price only as one column in that scorecard; the decision should reflect the full interaction cost and whether the system completes the work your product requires.

## Reader takeaway
MAI Realtime 1 is a new speech-to-speech option from Microsoft, with pricing beginning at $0.01 per audio minute and company-reported performance advantages over GPT Realtime 2.1. Put it in a controlled evaluation rather than treating the announcement as a production verdict. The winning voice model is the one that delivers acceptable latency, quality, and completed tasks at the actual cost of your workflow.

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Pricing](https://newruntime.com/topics/pricing.md): Explore the pricing topic hub.
- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)
