A generated title card for MAI Realtime 1 reading 'A cent a minute', with cards for the $0.01 audio-minute rate, the $1/$5 per-million reasoning tier, an unaudited 98.8% on Big Bench Audio, and 'Copilot Voice Mode: not announced', plus an OrcaRouter logo.
Guides & Insights

Microsoft's MAI Realtime 1 Lands at a Cent a Minute — and the Copilot Claim Isn't in the Paperwork

Author

jinhao.song

Date Published

Latest models · 20View all models →
Benchmarks: Artificial Analysis · updated daily
Back to all posts

On 17 November 2026, at Ignite, Microsoft announced seven models. One of them is MAI Realtime 1, the company's first speech-to-speech model, priced at $0.01 per audio minute. The X post that put it on our radar goes further than Microsoft's own pages do: it says Realtime 1 "will power Voice Mode in Copilot," and that the model "felt very close to the G​PT Live model." Neither claim appears anywhere in Microsoft's launch material. What the material says instead is narrower, more interesting, and easier to plan around: a limited-access preview, an O​penAI Realtime API-compatible interface, and a two-tier design that hands hard turns to a separate reasoning engine billed by the token.

That gap between the post and the paperwork is the story worth writing down on day one, because the second-hand version is already travelling further than the first-hand one. Here is what is documented, what is claimed without numbers, and what you can actually call today.

Two models shipped, not one

The post names both, and it is right to. The voice model is MAI Realtime 1. The transcription model is Dragon Copilot Transcribe, built on MAI Transcribe 2 — Microsoft's clinical dictation path, which the company says handles 100% of doctor–patient transcription inside Dragon Copilot, covers 10 million patient encounters a month, is used by a third of US physicians, and halves multilingual error rates. Those are Microsoft's figures, unaudited, and they describe a product line with years of deployment behind it rather than a model announced yesterday. It is the more mature half of the pair and the less discussed one.

MAI Realtime 1 is the new thing. Microsoft describes it as a speech-to-speech model "with natural conversational flow and precise steerability," combining audio understanding and voice generation with real-time reasoning so that a single interface can carry a steerable voice agent.

A generated scoreboard titled 'MAI Realtime 1 — the scoreboard' with five rows: Big Bench Audio 98.8% vendor, unaudited; voice price $0.01 per audio minute; reasoning tier $1.00 in and $5.00 out per 1M tokens; tau-Voice agentic no score yet; time to first audio P50 only. The footer reads 'Microsoft model page and rate card; the 98.8% is vendor-reported.'

What a cent a minute actually buys

The headline rate is $0.01 per audio minute of speech-to-speech. That is the number in Microsoft's own article title, and it is a genuine order-of-magnitude move against the voice APIs it is aimed at. But it is not the whole bill, and the structure matters more than the headline:

• Base conversation — $0.01 per audio minute, covering the speech-to-speech loop itself.

• Reasoning — billed separately at $1.00 per million input tokens and $5.00 per million output tokens.

• Preview billing — customers in the Foundry limited-access preview are not billed until public preview begins. Whatever you measure in the preview is free, which is the single best reason to start now rather than at general availability.

• What is not published — no figure for how often a conversation escalates into the reasoning tier. Without that, $0.01 is a floor and not an estimate. A support call that hits the reasoning engine on four turns costs meaningfully more than one that never leaves the fast path, and Microsoft has not said what the split looks like in practice.

Set that against the models it is being compared to on day one. O​penAI's G​PT Live 1 is $0.05 per minute for the voice session, billed by the second, with backend model and tool usage on top. O​penAI's G​PT Realtime 2.1 is priced per token instead — $32 per million audio input tokens and $64 per million audio output — which Artificial Analysis normalises to roughly $10.75 per hour at its High setting. ElevenLabs' Eleven Realtime Max lands around $4.80 per hour on the same independent basis. Microsoft's cent works out to $0.60 per hour before any reasoning tokens. On the voice line alone that is a real gap; once reasoning is in the bill it is a range, not a number.

A screenshot of Microsoft's MAI Realtime 1 model page captured 18 November 2026, showing the 98.8% Big Bench Audio accuracy figure, a 'Start a conversation' button, $0.01 per audio minute pricing, $5 per 1M output tokens, and a P50 time-to-first-audio card labelled 'P50 Latency on BigBench Audio'.

Four capability claims, two of them numbered

The capability list reads well and is entirely qualitative. Microsoft says the model handles interruptions and quick acknowledgements and knows when to let someone finish; that a fast system keeps dialogue real time while harder tasks route to a stronger reasoning engine without an awkward pause; that it still recognises requests over background noise and other voices; that it connects to enterprise systems to carry out multi-step requests; and that tone, style and behaviour can be steered through prompting.

Against that, the measurement side is thin and entirely self-reported. Microsoft says MAI Realtime 1 "sits on the performance-price Pareto frontier" for Big Bench Audio speech reasoning and the τ-Voice agentic task benchmark, and that it beats G​PT Realtime 2.1 on quality, speed and price. The one accuracy figure the company does publish is on the model page: 98.8% on Big Bench Audio, described there as "accuracy on spoken reasoning questions from the Big Bench Audio benchmark." That is a vendor number, produced by the vendor, on a benchmark the vendor chose, with no methodology note beside it. The same page carries a P50 "time to first audio" card and a second card labelled "P50 Latency on BigBench Audio," and no τ-Voice score of any kind. No language list is published either.

So the honest reading is narrower than either extreme. This is not a model with no numbers — it is a model with one accuracy number, one internal latency number, and no independent verification of either. Artificial Analysis' speech-to-speech board does not carry MAI Realtime 1 as of 18 November, which means nobody has re-run that 98.8% under a harness you can inspect. Treat it as a claim that happens to be quantified rather than as a result.

This is not a reason to dismiss the model. It is a reason to treat every comparison you read this week, including the one in the post that surfaced it, as an impression. "It felt very close to the G​PT Live model" is a testable statement that nobody has tested in public yet, and the independent harnesses have not scored MAI Realtime 1 either — Artificial Analysis' speech-to-speech board does not carry it as of 18 November. The comparison target Microsoft chose for itself is G​PT Realtime 2.1, a per-token reasoning voice model, not G​PT Live 1, the full-duplex one. Those are different products with different price structures, and the vendor head-to-head does not exist for the second.

The Copilot question

What Microsoft's materials actually say about Copilot is specific and much narrower than "Voice Mode." MAI Realtime 1 is coming to Dynamics 365 Contact Center for customer service, and it is coming to Dragon Copilot for simulated patient conversations that help nurses rehearse clinical interactions. Both are named, both are enterprise deployments, both are future tense.

Neither is Copilot Voice Mode. The Ignite keynote transcript does not mention Realtime 1 at all — it appears only in the related-stories list — and the model page does not mention Copilot in any form. So the claim that Realtime 1 will power Voice Mode in Copilot is, as of today, unsourced. It may well turn out to be true; consumer Copilot voice is an obvious place to put a model priced at a cent a minute. But a product decision that has not been announced is not the same as one that has, and planning a migration around an unannounced roadmap is how teams end up with an integration nobody supports.

A screenshot of Microsoft's 17 November 2026 launch post headed 'MAI Realtime 1: Leading voice AI at $0.01 per audio minute', tagged Ignite and Models, showing the announcement body and a Pareto frontier chart.

Who can call it, and how

Availability today is two doors, both narrow. There is a demo in MAI Labs, Microsoft's limited-preview surface for new and experimental models: two-minute sessions, sign-in required, 18 and over, personal non-commercial use only. During our check the demo itself reported "MAI Realtime is unavailable right now," which is worth knowing before you plan a team evaluation around it. The second door is the Foundry limited-access preview, reached through a registration form rather than a self-serve key.

The detail that matters most for engineering is compatibility: MAI Realtime 1 speaks the O​penAI Realtime API. If your voice stack is already built against that interface, an evaluation is a base-URL and model-string change rather than a rewrite — the same practical advantage ElevenLabs bought itself by shipping an OpenAI-compatible Realtime endpoint. If your stack is built against a different vendor's session protocol, the preview is a real porting project regardless of price.

The reasoning handoff is the part worth designing around

The two-tier design is the most consequential thing in the announcement, and it is the one the price headline hides. A cent buys the fast path. The moment a caller asks something that needs planning, tool calls across enterprise systems, or a careful multi-step answer, the turn leaves the speech model and lands on a reasoning model billed at $1 and $5 per million tokens. That is an ordinary text-model rate card attached to a voice product, which means your voice bill now has a text bill inside it and your latency budget has a second model in it.

That is precisely the seam a router belongs in. OrcaRouter passes provider list price through with 0% markup, so when a reasoning tier's rate changes the number your gateway reports changes the same day rather than at the next contract renewal. Automatic failover matters more here than in a chat product: a voice turn that stalls is a person listening to silence, and a request that falls back to a second provider in 400 milliseconds is invisible while one that retries for four seconds is not. The routing DSL is the other half — the fast conversational path and the reasoning escalation can be composed into one call so the handoff is a route rather than application logic you maintain yourself.

To be exact about what we do and do not carry: MAI Realtime 1 is not in the OrcaRouter catalogue. It is reachable through Microsoft's MAI Labs demo and the Foundry limited-access preview. The reasoning models behind a voice agent are a different matter — that is the layer we route, and it is the layer where a same-day price change or a provider hiccup actually reaches your users.

What to measure in the preview, since Microsoft measured it privately

Because the preview is unbilled until public preview, the cheapest useful work this month is measurement rather than procurement. Four numbers decide whether the cent holds up:

• Escalation rate — what fraction of turns leave the fast path for the reasoning tier. This converts $0.01 into an actual per-call cost, and it is the one figure Microsoft has left out entirely.

• Time to first audio, measured on your own audio path. Microsoft's model page shows a P50 figure for it, but a P50 with no distribution behind it tells you nothing about your tail, and the tail is what callers notice. Measure the slowest turn in a hundred, not the median.

• Interruption behaviour under real overlap — the vendor claim is that it distinguishes backchannels from interruptions from corrections. That is a claim you can falsify in an afternoon with recorded call audio, and it is the difference between an agent that feels natural and one that talks over people.

• Error rate on your own domain vocabulary. Contact-centre and clinical audio is where generic speech models fail, and Dragon Copilot's clinical numbers say nothing about your product names.

Run those against one incumbent rather than against a spec sheet. The honest read on 18 November is that MAI Realtime 1 has the most aggressive price in the category and the thinnest public evidence — one unaudited 98.8% and a latency percentile, against competitors with harnesses you can inspect — that its Copilot ambitions are documented only as contact-centre and nurse-training deployments, and that the claim it will power Copilot Voice Mode is a post, not a press release. A cent a minute is cheap enough to test and too cheap to trust unmeasured.