---
title: "MAI Realtime 1 enters preview: voice benchmarks, pricing and access"
description: "MAI Realtime 1 costs $0.01 per audio minute plus reasoning. Compare two voice benchmarks and understand personal-demo versus enterprise-preview access."
url: "https://promptszone.com/mai-realtime-1-preview-benchmarks-pricing-access"
published: "2026-11-18T19:00:05Z"
categories: ["New Models"]
---

# MAI Realtime 1 enters preview: voice benchmarks, pricing and access

Compare MAI Realtime 1 on two voice benchmarks, separate audio and reasoning charges, and check the limits of MAI Labs versus Foundry preview access.

![Selected voice benchmark comparison: MAI Realtime 1 leads Big Bench Audio at 98.9%; Gemini 3.8 Live Extended Thinking High leads the displayed τ-Voice scores at 68.6%. MAI figures are internal, peer figures attributed by Microsoft to Artificial Analysis.](https://promptszone.com/assets/uploads/e4e324cbdc9cbefc7e001990f11705d7.png)

**Summary:**

Microsoft's voice model leads spoken reasoning in a selected comparison, but not task completion. Preview access and separate reasoning charges matter.

Microsoft introduced MAI Realtime 1 on November 17, bringing live speech-to-speech conversations to its model lineup. The immediate choice is between a personal demo in MAI Labs and a limited-access enterprise preview in Microsoft Foundry, not a generally available production endpoint.

## Two benchmarks, different winners

The current [model page](https://microsoft.ai/models/mai-realtime-1/) combines Microsoft's internal MAI results with competitor figures attributed to Artificial Analysis. These are different evaluation origins, not an independently reproduced head-to-head. Among the four configurations below, MAI leads spoken reasoning; Gemini leads customer-service task completion.

| Benchmark (%) ↑ | MAI Realtime 1 | Gemini 3.8 Live Extended Thinking (High) | GPT-Realtime-2.1 High | Eleven Realtime v1 Max |
| --- | --- | --- | --- | --- |
| Big Bench Audio | **98.9** | 97.7 | 95.8 | 98.8 |
| τ-Voice | 57.5 | **68.6** | 45.7 | 47.6 |

Higher is better; bold marks the best displayed value. This is a selected comparison, not the entire field: Microsoft's τ-Voice chart also lists Phonic 1 (High) at 79.9%.

## The headline rate is not the whole bill

The [launch announcement](https://microsoft.ai/news/introducing-mai-realtime-1/) lists $0.01 per audio minute, with reasoning charged separately at $1 per million input tokens and $5 per million output tokens. Microsoft says limited-preview customers will not be billed until public preview begins.

For planning, 1,000 audio minutes would therefore have a $10 base component. A hypothetical additional one million reasoning input tokens and 200,000 output tokens would add $2, for $12 altogether. That is arithmetic using the announced rates, not a typical-call estimate or a preview invoice.

## Choose the right access route

MAI Labs requires sign-in and is restricted to adults and personal, noncommercial use, excluding school and work. An enterprise pilot belongs in the Foundry registration route. Microsoft describes OpenAI Realtime API compatibility, interruption handling and connections to enterprise tools; it names Agoda and Reservations.ai as early testers.

A useful pilot should measure successful task completion alongside turn-taking errors, latency and reasoning consumption. Include interruptions and noisy audio in the test set, and keep consequential tool actions behind explicit authorization. A strong spoken-answer score does not by itself establish reliable customer-service execution.

For the broader rollout, see our separate [MAI Code 2 and Flash comparison](https://promptszone.com/mai-code-2-flash-copilot-benchmarks-access).

## Sources and credits

Sources: Microsoft's November 17 launch announcement and current model-page benchmark data, checked November 18, 2026. Chart: PromptsZone. The official Microsoft provider logo is used as a fallback because no distinct model logo was verified, under [News Center editorial permission](https://news.microsoft.com/en-sg/default/). Used with permission from Microsoft. Microsoft does not endorse this article.
