Microsoft launches MAI Realtime 1 voice model at $0.01 per audio minute

Microsoft has introduced MAI Realtime 1, the first speech-to-speech model from its Microsoft AI (MAI) group, and priced it at $0.01 per audio minute for conversational use, with reasoning billed separately at $1 per million input tokens and $5 per million output tokens. The model was announced on November 17, 2026, during Microsoft Ignite in San Francisco, as one of seven MAI models presented at the conference, and Microsoft positions it for voice agents in enterprise settings such as customer service lines and booking assistants.
The launch is a preview. Users can sign in to MAI Labs, the company's public testing site, and request access to try the model, while businesses that want to build on it must register for a limited-access preview in Microsoft Foundry, the Azure platform where Microsoft sells its own and third-party models. Microsoft says preview customers will not be billed until a public preview begins, and it has not published a date for that stage or for general availability.
Microsoft claims MAI Realtime 1 outperforms OpenAI's GPT Realtime 2.1 on quality, speed and price across public benchmarks and its own product evaluations, naming Big Bench Audio for spoken reasoning, τ-Voice for agentic task completion and internal Dynamics 365 Contact Center tests. The announcement gives no scores for any of those benchmarks, and the model page currently shows placeholder values for latency and accuracy, so the comparison rests on Microsoft's description. Two customers, the travel platform Agoda and Reservations.ai, are named as early testers.
··········
MAI REALTIME 1 PRICING AND HOW THE BILLING WORKS.
The headline rate covers the speech-to-speech conversation itself. Microsoft describes MAI Realtime 1 as a two-part system in which a fast model keeps the dialogue flowing and hands harder requests to a stronger reasoning engine, and that reasoning step is metered in tokens on top of the per-minute charge. The official pages do not quote separate prices for audio input and audio output, which suggests the minute rate covers both directions of the call, but Microsoft has not spelled this out.
........
MAI Realtime 1 item | Official price | Unit |
Speech-to-speech conversation | $0.01 | Per audio minute |
Reasoning, input | $1.00 | Per 1M tokens |
Reasoning, output | $5.00 | Per 1M tokens |
Foundry limited-access preview | $0 | Not billed until public preview |
........
Prices as listed on Microsoft's official MAI Realtime 1 announcement and model page on November 18, 2026. Microsoft has not said whether the minute rate is rounded per call or billed per second, and it has not published regional pricing, discounts for committed volume or a price for the period after the preview ends.
··········
MAI REALTIME 1 VS GPT-LIVE-1 AND GPT-REALTIME-2.1 ON PRICE.
The closest like-for-like comparison is OpenAI's GPT-Live-1, a full-duplex voice model that OpenAI also bills by the minute. Its official model page lists voice sessions at $0.05 per minute, billed per second without rounding up to the next minute, with backend model and tool calls charged separately at the normal rates of the model a developer configures. On the headline rate alone, MAI Realtime 1 costs one fifth as much per minute of conversation.
GPT Realtime 2.1, the model Microsoft names in its benchmark claims, is priced in tokens, so its cost per minute depends on how much audio and text a session produces and cannot be stated as a single figure. The table lists each model in the unit its provider uses, without conversion.
........
Model | Provider | Conversation price | Reasoning or backend | Status |
MAI Realtime 1 | Microsoft | $0.01 per audio minute | $1 input / $5 output per 1M tokens | Limited-access preview |
GPT-Live-1 | OpenAI | $0.05 per minute, billed per second | Billed at the configured model's rates | Available in the API |
GPT-Realtime-2.1 | OpenAI | Audio $32 input / $0.40 cached / $64 output per 1M tokens | Text $4 input / $0.40 cached / $24 output per 1M tokens | Available in the API |
GPT-Realtime-2.1 mini | OpenAI | Audio $10 input / $0.30 cached / $20 output per 1M tokens | Text $0.60 input / $0.06 cached / $2.40 output per 1M tokens | Available in the API |
........
OpenAI prices as listed on the official GPT-Live-1, GPT-Realtime-2.1 and GPT-Realtime-2.1 mini model pages on November 18, 2026. The comparison is not fully like-for-like, because MAI Realtime 1 is still in a preview with no billing, while the OpenAI models are generally available with published rate limits. GPT-Live-1, for example, is limited by concurrent sessions, from 50 on OpenAI's Build tier to 500 on Grow, and is not offered on the free tier. Microsoft has not published equivalent limits for its model. Our earlier coverage of OpenAI's GPT-Live-1 API launch and its $0.05 per minute pricing describes how that model splits voice and backend reasoning.
··········
WHAT A CONTACT CENTER WORKLOAD WOULD COST ON MAI REALTIME 1.
To show how the two billing components combine, consider an illustrative support line handling 5,000 calls a day with an average length of four minutes, which adds up to 20,000 audio minutes daily. Assume that 30% of calls need the reasoning engine, for example to check a booking or apply a refund rule, and that each of those escalations consumes 4,000 input tokens and 600 output tokens. These assumptions are ours and are meant only to make the arithmetic concrete.
........
Daily cost item | MAI Realtime 1 | GPT-Live-1 |
Audio minutes per day | 20,000 | 20,000 |
Conversation charge | $200.00 | $1,000.00 |
Reasoning, 1,500 escalated calls | $10.50 | Depends on chosen backend model |
Daily total before backend costs | $210.50 | $1,000.00 |
30-day total before backend costs | $6,315 | $30,000 |
........
On MAI Realtime 1, the conversation charge is 20,000 minutes multiplied by $0.01, or $200 a day. The reasoning charge comes from 6 million input tokens at $1 per million ($6) and 900,000 output tokens at $5 per million ($4.50). In this scenario the per-minute fee accounts for about 95% of the bill, so call length matters far more than reasoning volume. For GPT-Live-1 the session fee alone is $1,000 a day, and the backend model's token costs would come on top, which is why the table stops at the session charge for OpenAI.
The calculation favors high-volume, long-duration voice workloads such as reservations, order status lines and appointment scheduling, where minutes accumulate quickly and most turns are routine. Deployments that send nearly every turn through the reasoning engine would see the token component grow, although at $5 per million output tokens it would take very long answers on most calls to rival the minute fee. These figures will only become real costs once Microsoft starts billing, and the company could change the price before then.
··········
MAI REALTIME 1 AVAILABILITY IN MAI LABS AND MICROSOFT FOUNDRY.
There are two ways to reach the model today, and they serve different purposes. MAI Labs, at labs.microsoft.ai, lets signed-in users who complete an interest form hold a conversation with MAI Realtime 1, but its terms state that MAI Labs is intended for personal, non-commercial use only, so it works as a demo for evaluating voice quality and turn-taking. Companies that want to put the model into a product need to fill in the interest form for the Microsoft Foundry limited-access preview, where Microsoft decides who gets in and has not described the selection criteria.
For developers, the most practical detail is compatibility. Microsoft says MAI Realtime 1 works with the OpenAI Realtime API, which means a team that already runs a voice agent on GPT Realtime models should be able to test the Microsoft model by changing the endpoint and model settings rather than rewriting its audio pipeline. The company has not yet published a model card, supported language list, voice catalog, context window or rate limits for MAI Realtime 1, so teams evaluating it should plan their tests around those gaps.
Inside Microsoft's own products, the model is planned for Dynamics 365 Contact Center, where it would power voice agents for customer service, and for Dragon Copilot, where it would run simulated patient conversations that help nurses practice clinical interactions. Neither integration has a release date. The new model joins the speech components Microsoft released earlier in the autumn, covered in our article on MAI-Transcribe-2-Streaming and MAI Voice 2.1 for real-time multilingual agents, which handled transcription and speech output as separate steps.
··········
HOW MAI REALTIME 1 HANDLES INTERRUPTIONS, NOISE AND ACTIONS.
Microsoft describes several conversational behaviors that matter for phone and app agents. The model is designed to cope with interruptions, to give short acknowledgments while a caller is still speaking, and to judge when a speaker has finished rather than cutting in during a pause. It is also meant to pick out the caller's request when there is background noise or more than one person talking, a common failure point for voice systems used in cars, shops and call centers.
The split between a fast conversational layer and a deeper reasoning engine is Microsoft's answer to the latency problem of voice agents, where a model that stops to think produces long silences. According to the company, the hand-off happens without awkward pauses, although it has not published latency measurements to support that claim. The model can connect to enterprise systems and tools to carry out requests, and its tone, style and persona can be adjusted through prompting for different brands and use cases. Microsoft's pages do not use the term function calling or document the tool interface, so the exact mechanics will only be clear once the Foundry documentation appears.
··········
THE OTHER MICROSOFT AI MODELS ANNOUNCED AT IGNITE 2026.
MAI Realtime 1 was presented alongside six other models. Microsoft already detailed its coding models in GitHub Copilot, which we covered in our report on MAI Code 2 and MAI Code 2 Flash. The remaining announcements cover images, security, healthcare transcription and model routing.
........
Model | Use | Availability | Official price |
MAI Image 2.7 | Image generation and editing | Public preview in Foundry, MAI Labs | About $35 per 1,000 images |
MAI Image 2.7 Flash | High-volume image generation | Public preview in Foundry, MAI Labs | About $10 per 1,000 images |
MAI Cyber 2 Flash | Vulnerability discovery and remediation | Microsoft security products, Codename MDASH from mid-December | List price $0.10 input / $0.50 output per 1M tokens |
Dragon Copilot Transcribe | Clinical transcription on MAI-Transcribe-2 | Live in Dragon Copilot | Not published |
Microsoft-Decision-1.1 | Routing requests between models | Generally available in Foundry, OpenRouter, Vercel | Same price as Decision-1, no figure given |
........
Microsoft says MAI Image 2.7 ranks first on the Artificial Analysis image-editing leaderboard with 1,215 Elo across 5,478 votes as of November 16, four points ahead of OpenAI's GPT Image 2.6 Sunburst at its maximum setting, and it is now the default model family in Bing Image Creator. For MAI Cyber 2 Flash, the company reports 92.3% on CyberGym-E2E against 91.5% for GPT-6 Astra, after excluding 92 of the benchmark's 905 tasks that overlapped with its training data, and it says comparisons with GPT-6 Cyber will follow in the coming weeks.
··········
MAI REALTIME 1: CONFIRMED FACTS AND OPEN QUESTIONS.
The confirmed facts are the price structure, the preview status and the access routes. MAI Realtime 1 costs $0.01 per audio minute plus $1 and $5 per million reasoning tokens, it can be tried for personal use in MAI Labs, business access goes through a Foundry limited-access preview with no charges until public preview, and the model accepts requests in the OpenAI Realtime API format. Microsoft has also committed to using it in Dynamics 365 Contact Center and Dragon Copilot.
The performance claims are less settled. Microsoft says the model beats GPT Realtime 2.1 on quality, speed and price, but it has published no benchmark scores, no latency figures and no language list, and its own disclaimer notes that results vary by application and methodology. Until Microsoft releases a model card and opens the public preview, the price advantage over GPT-Live-1 is the clearest established difference: on the published rates, an hour of conversation costs $0.60 on MAI Realtime 1 and $3.00 in GPT-Live-1 session fees.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org




