Microsoft announced MAI Realtime 1 on November 17, 2026, introducing its first speech-to-speech model for live voice applications. The model starts at $0.01 per audio minute, with reasoning billed separately, and Microsoft says it is compatible with the OpenAI Realtime API. Developers can try it in MAI Labs or register for a limited-access Microsoft Foundry preview. Microsoft’s announcement says Foundry preview customers will not be billed until public preview begins.
Those details make this a useful release for teams already building voice agents. API compatibility offers a potential shortcut to evaluating another model without redesigning the entire application. The separate reasoning charges, however, mean the headline audio rate is only one component of the eventual bill.
The access distinction matters just as much: MAI Labs is a short, personal-use demonstration; Foundry is the registration-based route for enterprise evaluation. Treating them as interchangeable would lead to the wrong testing plan.
What Microsoft’s First Speech-to-Speech Model Does
MAI Realtime 1 accepts spoken interaction and produces spoken responses. Microsoft’s model page describes it as combining audio understanding, a natural voice and real-time reasoning through one interface.
That changes what developers integrate. A conventional voice-agent pipeline can connect speech recognition, a text-based LLM and text-to-speech as separate services. A speech-to-speech interface packages the conversational audio interaction behind a single model interface. Developers still need application logic, permissions and connections to business systems, but they have fewer separate speech components to coordinate.
According to Voice AI Space’s account of the launch, Microsoft describes five main capabilities:
- Turn-taking: handling interruptions, brief acknowledgments and the end of a speaker’s turn.
- Reasoning on demand: using a fast conversational system that can invoke stronger reasoning when needed.
- Noise handling: recognizing requests amid background noise or other speakers.
- Tool use: connecting conversations to enterprise systems that can take actions.
- Prompted personas: setting tone, style and behavior through instructions.
These are consequential capabilities for a voice agent. An appointment assistant, for example, needs to distinguish a brief acknowledgment from a new request, wait for availability data and avoid talking over a correction. A fluent voice alone cannot solve that interaction.
My view is that turn-taking and successful tool use deserve at least as much attention as voice quality during evaluation. They determine whether a conversation actually gets something done.
Pricing Separates Audio From Reasoning
Microsoft lists three billing components for MAI Realtime 1. The audio rate covers base speech-to-speech usage, while reasoning has separate input and output token prices. The model page confirms the following rates:
| Billing component | Published price | Budgeting implication |
|---|---|---|
| Base speech-to-speech | $0.01 per audio minute | Audio usage creates the baseline charge |
| Reasoning input | $1 per million input tokens | Reasoning input adds to the audio bill |
| Reasoning output | $5 per million output tokens | Reasoning output costs five times as much per token as input |
A token is a unit of model processing, not an audio minute. The two meters therefore cannot be converted into a single reliable per-call price without knowing both the audio usage and the reasoning workload.
Using Microsoft’s listed rates, the cost calculation is:
Audio minutes × $0.01 + reasoning input tokens ÷ 1,000,000 × $1 + reasoning output tokens ÷ 1,000,000 × $5.
For an illustrative workload with 1,000 billable audio minutes, 2 million reasoning input tokens and 500,000 reasoning output tokens, the calculation would be:
| Component | Illustrative usage | Calculated charge |
|---|---|---|
| Audio | 1,000 billable minutes | $10.00 |
| Reasoning input | 2 million tokens | $2.00 |
| Reasoning output | 500,000 tokens | $2.50 |
| Total for those components | $14.50 |
This is a budgeting example, not a measured workload or a prediction of typical usage. It shows why quoting only the audio rate can understate costs: in this example, reasoning adds 45% to the base audio charge.
The on-demand reasoning design also makes conversation complexity economically relevant. If one workflow invokes more reasoning than another, the two can have different costs despite using the same number of billable audio minutes. A short but complicated request could have a different cost profile from a longer, routine exchange.
An Audio Minute Still Needs a Definition
Microsoft’s published description leaves an important billing question unresolved: whether the per-minute rate counts incoming audio, outgoing audio or both. Voice AI Space explicitly flags that omission.
At the listed rate, 60 billable audio minutes would cost $0.60 before reasoning. That arithmetic should not be presented as a guaranteed price for a one-hour call, because elapsed call time and metered audio time may differ.
I’d ask for the precise metering rules before comparing MAI Realtime 1 with an existing provider. A fair comparison needs the same conversation workload, a clear accounting of both audio directions and all reasoning charges. Otherwise, a cheap-looking per-minute figure can hide differences in what is being counted.
OpenAI Realtime API Compatibility Helps Evaluation
Microsoft says MAI Realtime 1 is compatible with the OpenAI Realtime API. For teams already using that interface, this could be the most immediately useful part of the announcement.
API compatibility concerns how an application communicates with a service. It can reduce integration work by allowing developers to retain familiar interaction patterns while evaluating a different model. It does not make the models identical in behavior.
The practical opportunity is to separate two decisions: changing the application’s interface and changing its model provider. If the existing integration can be adapted with limited changes, a team can spend more of its evaluation effort on conversation quality, responsiveness and cost.
A migration assessment should still check:
- Connection setup: endpoints, authentication and supported connection methods.
- Session configuration: audio settings, instructions and available options.
- Conversation events: how the application receives speech, response and interruption events.
- Tool interactions: argument generation, result handling and recovery from failed actions.
- Operational behavior: timeouts, reconnects, usage reporting and service limits.
These are validation points, not a list of documented incompatibilities. The launch evidence establishes Microsoft’s compatibility claim, but it does not provide a detailed feature-by-feature migration matrix.
A familiar API also cannot guarantee that an existing prompt will produce the same results. Different models can interpret instructions differently, choose tools at different moments or handle interruptions differently. Successful connection is the first milestone; successful completion of the intended task is the meaningful one.
I’d use compatibility as a reason to run a controlled comparison, not as a reason to skip one. Reusing the same scenarios and success criteria makes the evaluation more useful than comparing unrelated demos.
MAI Labs and Foundry Serve Different Purposes
Microsoft offers two ways to approach MAI Realtime 1, with different conditions attached.
| Access route | What is available | Important condition |
|---|---|---|
| MAI Labs | An interactive voice demonstration | Personal use only, with sessions up to two minutes |
| Microsoft Foundry | A limited-access model preview | Registration is required; preview customers are not billed until public preview begins |
The MAI Labs page labels the demonstration “For personal use only” and “Not for commercial, school or work use.” It also states that sessions can last up to two minutes. The interface requires sign-in and microphone access.
Those conditions make Labs useful for a brief, personal introduction to the interaction style. They also put a clear boundary around its role. A two-minute demo cannot establish how a production agent behaves across long conversations, repeated tool calls or sustained traffic, and its stated terms rule out treating it as a work-use sandbox.
The Foundry limited preview is the relevant registration path for teams pursuing an enterprise integration. Microsoft’s announcement says participating preview customers will not be charged until public preview starts. That is a billing condition tied to the preview phase, not a promise of permanently free usage.
The distinction matters for scheduling as well as cost. A model being announced does not mean every developer can immediately provision production access. Limited preview requires registration, so an evaluation plan needs to account for getting access before implementation work can proceed.
Frequently Asked Questions
4 questions
1Is MAI Realtime 1 Microsoft’s First Speech-to-Speech Model?
Yes. Microsoft introduced it on November 17, 2026, as its first speech-to-speech model, combining live audio interaction with real-time reasoning.