98.9. That's the score Microsoft reports for MAI Realtime 1 on Big Bench Audio, a test where a voice model has to answer reasoning questions asked out loud. It's the top score on the chart Microsoft published on Tuesday. A little further down the same page there's another number: 57.5. That's the share of customer-service tasks the same model completes on τ-Voice, a test where understanding the caller isn't enough. The model has to get the job done.
Same model, same page. Side by side, those two numbers say more about where voice AI stands than any launch slogan. If you sell over the phone, they're the most useful thing to come out of this week.
What Microsoft shipped
On November 17, at its Ignite conference, Microsoft AI announced seven new models. Only one of them is about the phone, and it's the first of its kind Microsoft has built in-house: MAI Realtime 1, a speech-to-speech model. It hears the caller and answers out loud, with no written transcript in between. Google and OpenAI already have models like this. Now Microsoft does too.
It works on two levels. A fast layer keeps the conversation moving in real time. When a request gets complicated, it calls on a heavier reasoning engine, and Microsoft says it does that without awkward pauses. Median time to first audio, the gap between the end of a question and the start of the answer, is 880 milliseconds. Under a second.
Microsoft is also clear about where it wants the model to work: inside Dynamics 365 Contact Center, its call-center product, and in Dragon Copilot, where it plays simulated patients so nurses can practice. Among the first companies testing it, Microsoft names Agoda, the travel booking platform.
How it behaves on a call
The feature list is short and goes straight for the problems anyone who has fought an automated phone line will recognize. The model handles interruptions. It picks up the little "uh-huh"s and "okay"s we drop while the other person is talking, and doesn't mistake them for an answer. It knows when to wait for someone to finish. It catches the request even with background noise or other people talking nearby. It connects to business systems, so it can do things instead of just answering. And its persona and tone are set with instructions, not months of training.
The model page has six sample calls, each with a different personality. In one, it's the front desk at a dental practice: it finds a time that suits the caller, leaves the hygienist a note about the patient's sensitive gums and confirms the booking. In another it works for a credit card company. When the caller gets distracted it repeats the question, when asked to keep it short it does, and it checks before closing the account. In a third it changes a flight, and when the later flight is over the traveler's budget, it finds a better option the next morning.
These are demos picked by the people selling the model, of course. But they make the point. The voice isn't the problem anymore. The manners and the patience are already there. What matters now is what happens during the call: the note for a colleague, the slot that has to fit, the confirmation before anything gets changed.
The numbers
Microsoft published two charts. We've kept the models from the major labs. The first is Big Bench Audio: reasoning questions asked by voice, scored on how many answers are right.
| Model | Big Bench Audio |
|---|---|
| MAI Realtime 1 (Microsoft) | 98.9% |
| Gemini 3.8 Live Extended Thinking, High (Google) | 97.7% |
| Grok Voice 2.0 (xAI) | 97.2% |
| GPT-Realtime-2, High (OpenAI) | 96.6% |
| GPT-Realtime-2.1, High (OpenAI) | 95.8% |
| Gemini 3.8 Live (Google) | 91.7% |
| GPT-Live-1 Astra, medium (OpenAI) | 90.1% |
| GPT-Live-1 Sol, low (OpenAI) | 89.0% |
| GPT-Realtime-2.1 Mini, High (OpenAI) | 74.7% |
Source: Microsoft AI, MAI Realtime 1 model page, "Big Bench Audio score" chart, November 17, 2026. Other models' scores measured by Artificial Analysis; MAI Realtime 1 from Microsoft's internal evaluation. Higher is better.
The second is τ-Voice. The model plays a customer-service agent and has to complete real tasks, with policies to follow and systems to use, while talking to a simulated customer. What counts is how many tasks it gets right.
| Model | τ-Voice, tasks completed |
|---|---|
| GPT-Live-1 Astra, medium (OpenAI) | 74.5% |
| Grok Voice 3.0 (xAI) | 71.6% |
| Gemini 3.8 Live Extended Thinking, High (Google) | 68.6% |
| GPT-Live-1 Sol, low (OpenAI) | 59.3% |
| MAI Realtime 1 (Microsoft) | 57.5% |
| Qwen Audio 3.0 Realtime Plus (Alibaba) | 54.6% |
| GPT-Realtime-2.1, High (OpenAI) | 45.7% |
| Qwen Audio 3.0 Realtime Flash (Alibaba) | 35.9% |
| Gemini 3.8 Live (Google) | 30.1% |
| GPT-Realtime-2.1 Mini, High (OpenAI) | 29.4% |
Source: as above, "τ-Voice task completion" chart. Full-duplex models only, meaning they listen and speak at the same time. The original chart also includes models from two specialist vendors that we've left out; one of them tops this test. Higher is better.
Understanding isn't doing
Put the two tables next to each other and the most interesting thing in the announcement jumps out, even though Microsoft doesn't point to it. On Big Bench Audio, nearly every major model sits around or above 90%. Understanding a spoken question and reasoning about it is, for practical purposes, a solved problem. On τ-Voice, no major-lab model reaches 75%, and the same names are spread between 29 and 74.5.
MAI Realtime 1 is the clearest case: first on the first test, mid-table on the second. Gemini 3.8 Live, in its fast version, scores 91.7 on one and 30.1 on the other. Give it time to think, in the Extended Thinking version, and it climbs to 68.6.
For anyone who sells by phone, that's good news, and practical news. The part people laughed at two years ago, the AI that doesn't understand what you just said, now works. What's left to get right is different. You give the agent a clear task, hook it up to the right systems, and set firm rules on what it can do on its own and what it has to confirm. That's call design, not technology. And you can do it now.
Pick the model for the job
There's a second way to read the tables, and it's the more useful one: no model wins everything. The leader in spoken reasoning isn't the leader in finishing tasks, and the fast version of a model can behave very differently from the one that takes time to think. Counting Microsoft, at least five labs now ship a serious voice model (Google, OpenAI, xAI, Alibaba and Microsoft), most of them in several versions.
That changes how you choose. An appointment-confirmation call is short and needs snap. You want a fast model that never leaves the caller hanging. A call that has to check details, offer three time slots and write a note in the CRM needs a model that finishes tasks, even if it takes half a second longer. MAI Realtime 1 is compatible with OpenAI's Realtime API, which helps: if you already have an agent built that way, you can try it without starting over.
What the announcement leaves out
Three things, and the first one matters.
Languages. Neither the announcement nor the model page says which languages MAI Realtime 1 speaks. Microsoft's other voice models cover plenty of them, but for this one it isn't stated, and every published sample is in English. Until Microsoft says so, don't assume it.
Who measured what. The other models' scores come from Artificial Analysis, an independent benchmarking firm. MAI Realtime 1's come from Microsoft's own evaluation, and Microsoft itself notes that results depend on the evaluation setup and vary by application. Read the comparison as a guide, not an official ranking.
Where it stands. The model is in preview. You can try it in MAI Labs, in conversations of up to two minutes and for personal use only, or sign up for the limited-access preview on Foundry. Putting it on the phones every day will have to wait a little.
What to do Monday morning
Take the calls your business makes every week and split them into two columns. In the first, put the ones where understanding and answering well is enough: confirming an appointment, reminding someone of a deadline, giving information. In the second, put the ones where something has to happen at the end. A time booked in the calendar, a record updated in the CRM, a note for the colleague who's visiting the customer.
Say you install solar panels and make 300 calls a week. The reminder about tomorrow's site visit goes in column one. The call that has to find the surveyor a free Tuesday, ask what kind of roof it is and log it in the CRM goes in column two. Column two is usually the shorter one. It's also the one worth more.
The first column is already safe ground for a voice agent, with almost every model of this generation. The second is where models differ, and where a well-designed agent pulls away from one thrown together in a hurry. Next to each call in the second column, write down which system it has to touch: calendar, CRM, email. That list is where any serious voice agent project starts, and it'll take you an hour.
We work on two specific calls, and both sit in the second column: cold calls that book meetings and reactivating leads that went quiet.
Sources
- Microsoft AI, Introducing MAI Realtime 1, November 17, 2026 (capabilities, evaluation notes, availability, early customers)
- Microsoft AI, MAI Realtime 1 model page (Big Bench Audio and τ-Voice charts, time to first audio, sample calls)
- Microsoft AI, Seven new world class models, November 17, 2026 (Ignite announcements)
Want to hear it on your own calls? We build and run the agent for you. You hear it in your industry, approve how it talks, and get the meetings. → Book a call
