# Microsoft MAI Realtime 1: $0.01 Audio Minutes Plus Reasoning

> Source: <https://www.digitalapplied.com/blog/microsoft-mai-realtime-1-voice-ai-price>
> Published: 2026-11-19 00:00:00+00:00

**Key takeaway:** MAI Realtime 1 bills $0.01 per billable audio minute plus reasoning at $1 per million input tokens and $5 per million output tokens, and Microsoft's $1.3548-per-input-audio-hour benchmark figure reflects its evaluation workload, not connected phone time.

A five-minute support call may contain two minutes of customer speech, a minute of agent speech and two minutes of silence while a booking system responds. Microsoft prices MAI Realtime 1 at $0.01 per audio minute, plus separately metered reasoning. The launch announcement does not say how that five-minute call maps to its audio meter. For a contact center planning a pilot, the missing measurement is billable usage on a call that actually gets resolved.

Microsoft introduced its first speech-to-speech model on November 17. It accepts live audio, responds in speech and can call tools. The company is offering personal experimentation in MAI Labs and registration for a limited-access Microsoft Foundry preview. Microsoft says customers in that preview will start being billed when public preview begins. The announced rates are enough to build a forecast today. Production pricing evidence from a customer’s own traffic has to come later. [Microsoft’s launch and pricing footnotes](https://microsoft.ai/news/introducing-mai-realtime-1/)

1. 01Forecast audio and reasoning separatelyThe announced rates are $0.01 per billable audio minute, $1 per million reasoning input tokens and $5 per million reasoning output tokens. The limited-access preview is unbilled until public preview.
2. 02The published benchmark price uses an input-audio hourMicrosoft’s comparison plots MAI at $1.3548 per normalized hour of input audio. That figure belongs to the evaluation workload and does not establish what an hour of connected phone time costs.
3. 03OpenAI and xAI count different thingsGPT-Realtime-2.1 meters audio tokens by direction. GPT-Live-1 charges for open session time plus its backend. Grok Voice 3.0 changes its audio meter depending on the turn-detection mode.
4. 04Use the business preview for a company trialThe browser demo requires a personal, non-commercial use case. Foundry is the access route for an employer evaluating its own integration and customer-data requirements.

## 01 — Speech to speechWhich part of the call does *MAI handle?*

In a conventional phone agent, speech recognition produces text, a language model decides the reply and a separate system speaks it. MAI Realtime 1 exposes a conversational voice interface. Microsoft describes a fast voice layer that handles the exchange and can call on a deeper reasoning engine. The goal is to keep brief acknowledgments and interruptions responsive while allowing a longer decision when a caller asks about policy or changes their request.

Consider a guest moving a reservation. The voice model needs to hear a correction to the date while it is speaking. The reasoning engine may need to interpret the cancellation policy. A booking tool still owns the room availability, the fee and the actual change. A model that politely accepts the correction but commits the original date has failed the call. This is why an integration test needs the reservation record as well as the audio. [How full-duplex voice systems handle turn-taking](https://www.digitalapplied.com/blog/full-duplex-voice-models-compared-2026)

Microsoft reports support for background noise, prompted tones and enterprise tool calls. It names Agoda as an early tester and describes integration work for Dynamics 365 Contact Center and simulated patient conversations in Dragon Copilot. Those examples identify intended uses. The launch does not publish a call-resolution rate for that business or a production service commitment for those integrations.

## 02 — Billing unitsWhat goes into the *model bill?*

Use three numbers. Multiply the provider’s billable audio minutes by $0.01. Add reasoning input tokens divided by one million at $1, then reasoning output tokens divided by one million at $5. The two token rates apply to the deeper reasoning work. Reading them as the rates for the caller’s speech or for every word synthesized by the voice layer would mix two different meters. [Microsoft’s speech and reasoning rates](https://microsoft.ai/news/introducing-mai-realtime-1/)

Sixty *billable audio* minutes yield a $0.60 speech charge before reasoning. Microsoft’s public description does not specify how it treats silence, simultaneous speech, rounding or an open connection waiting for a tool. A carrier measures how long the line stays connected. Counting the duration of a recording, or adding the caller’s and agent’s speaking times, gives you useful operational data. Either becomes an estimate of Microsoft’s meter only after you establish the relationship with preview telemetry or the provider.

Keep telephone numbers, carrier minutes, call recording, storage, external tools and human escalations on their own lines. Some contact centers already pay for these under an existing agreement. Charging each model the same shared costs makes sense only if it uses them in the same way. An assistant that keeps a caller on the line through two slow lookups can increase the telephone charge even if its own model charge stays small. [A guide to voice AI billing units](https://www.digitalapplied.com/blog/voice-ai-pricing-characters-tokens-minutes)

## 03 — Worked budgetPrice *1,000 support calls*

Suppose a reservations team models 1,000 calls. Its planning assumptions are 3,600 billable audio minutes, 1.5 million reasoning input tokens and 300,000 reasoning output tokens across the batch. These are illustrative usage totals, independent of how long the telephone lines stay open. At the published rates the three charges are $36, $1.50 and $1.50. The projected model bill is $39, or 3.9 cents per attempted call.

If 800 of those calls end with the requested action correctly completed without staff help, the model charge works out to $39 divided by 800, or about 4.9 cents per automated resolution. That figure includes the model spend on the other 200 attempts. Their staff time and the cost of any corrections still belong in the business case. Count repeat calls about the same failed change so a large number of short calls cannot disguise a poor outcome.

Now keep all usage fixed except reasoning output, which rises from 300,000 to 900,000 tokens after the team introduces more complicated fare rules. The extra 600,000 tokens add $3. The forecast becomes $42. If the extra reasoning resolves 900 calls unaided, the model cost per automated resolution falls to about 4.7 cents. If it still resolves only 800, it rises to 5.25 cents. That is a decision the team can test without assuming that deeper reasoning either pays for itself or wastes money.

## 04 — Competing metersWhat would the *same five-minute call* cost elsewhere?

For a concrete comparison, return to a five-minute open session containing two minutes of user speech and one of assistant speech. The figures below isolate each provider’s stated meter as of November 19. The extra two minutes affect some of the calculations. A fair invoice comparison would also need the text and tool traffic, repeated context and each provider’s reported usage.

##### GPT-Realtime-2.1

Uncached audio costs $32 per million input tokens and $64 per million output tokens. At roughly 600 user tokens or 1,200 assistant tokens per minute of speech, the two incoming minutes cost $0.0384 and the outgoing minute costs $0.0768. That is $0.1152 for the new audio alone.

##### GPT-Live-1

The voice session costs $0.05 per minute, billed by the second through closure. Five minutes of an open session costs $0.25, including silence and backend wait time. WebRTC session creation bills 15 seconds up front, credited against running duration. Backend model and tool charges are additional.

##### Grok Voice 3.0

The rate is $0.08 per minute. Default server-side voice activity detection bills session duration, so five minutes costs $0.40. Push-to-talk bills audio sent and received. If the client actually sends two minutes and receives one, the audio charge is $0.24.

OpenAI’s Realtime guide converts the duration of a fresh utterance to audio tokens. It also describes a conversation whose earlier messages become input to later turns and may qualify for caching. That is why the $0.1152 fresh-audio example is neither a full invoice nor a prediction of how the stream handles interruptions. The comparable MAI figure remains unknown until you measure its audio and reasoning meters. [OpenAI API prices](https://developers.openai.com/api/docs/pricing) · [OpenAI voice cost guide](https://developers.openai.com/api/docs/guides/voice-latency-cost) · [xAI audio and text-event billing](https://docs.x.ai/developers/models/speech-to-speech) · [Grok Voice 3.0 and GPT-Live-1.1 availability](https://www.digitalapplied.com/blog/grok-voice-3-vs-gpt-live-1-1-voice-ai-compared)

## 05 — Performance evidenceWhat did Microsoft *actually measure?*

Microsoft reports 98.9% accuracy on Big Bench Audio, a spoken reasoning test, and 880 milliseconds median time to first audio on that benchmark. On τ-Voice, which exercises the agent’s ability to complete tasks in a spoken conversation, Microsoft reports 57.5% task completion. A correct answer to a spoken question and a completed service transaction have different success criteria. The latter is closer to the reservation team’s job. [Microsoft’s MAI Realtime 1 model page and performance charts](https://microsoft.ai/models/mai-realtime-1/)

On the same charts, Microsoft lists GPT-Realtime-2.1 at high reasoning effort with 95.8% on Big Bench Audio and 45.7% on τ-Voice. Its τ-Voice chart also lists Grok Voice 3.0 at 71.6%. Microsoft credits Artificial Analysis for the competitor points and an internal evaluation for its own. Its claim that MAI outperformed GPT-Realtime-2.1 on quality, speed and price refers to particular configurations and includes Dynamics 365 evaluations. It does not establish the result for every competing configuration or language.

Both charts place MAI at $1.3548 per hour of **input audio**. For Big Bench Audio, Microsoft normalizes the cost using a fixed subset of 40 questions. This describes the cost of running the workload scaled by the duration of its inputs, including work to answer those inputs. It uses a different denominator from the base speech tariff. The gap between $1.3548 and $0.60 does not reveal what share of a telephone invoice will be reasoning. Likewise, 880 ms is a benchmark median. A caller waiting for a live hotel system cares about the slow end of the distribution and about when they get the answer. [How voice benchmarks differ from listener preference](https://www.digitalapplied.com/blog/realtime-voice-models-benchmarks-vs-listener-preference)

## 06 — Preview termsWhich trial can *a business use?*

For personal exploration, the MAI Labs demo lets an adult sign in with a Microsoft account and use a microphone. Its notice limits the demo to personal, non-commercial use and excludes work and school activities. It asks for consent from anyone else whose voice you share and says to leave out sensitive information. A personal conversation can show how the assistant responds when you interrupt or pause. A team evaluating a new reservations line should register for the Foundry limited-access preview. Signing into the public demo does not grant that access. As of November 19, OpenRouter’s public model catalogue listed no MAI Realtime 1 model ID. [Model demo and Foundry interest form](https://microsoft.ai/models/mai-realtime-1/) · [MAI Labs terms](https://microsoft.ai/models/terms/) · [OpenRouter model catalogue](https://openrouter.ai/api/v1/models?output_modalities=all)

The model-specific Labs notice says audio is processed only for the live conversation. Microsoft’s general Labs FAQ separately describes an up-to-29-day period for uploaded files and generated content, and session-only handling for its distinct DuoAI experience. These statements cover different data and products. They do not establish a Foundry retention period for MAI Realtime recordings, transcripts or tool payloads. Before routing customer calls, obtain the preview terms for that deployment and check the contact center’s own recorder, transcript archive and support logs. Those systems can create their own copies. [Microsoft’s Labs storage FAQ](https://microsoft.ai/models/faq/)

## 07 — Replacement testTry one *booking change* end to end

Microsoft advertises compatibility with the OpenAI Realtime API. That gives an existing team a starting point for its client and event handling. Once admitted to Foundry, run a scripted, fictional booking against a test reservation system. Keep the same microphone recording or caller script, availability data, tool definitions and success criteria for the current model and MAI. Check the connection and audio format, interruption events, tool argument format, cancellation and recovery after a dropped connection before judging how the voice sounds.

Try a caller who asks for Friday, corrects it to Saturday halfway through the reply, then pauses while looking for the booking reference. Let the first availability lookup time out. A passing run stops the obsolete reply, keeps the corrected date, retries the lookup safely and confirms the exact fee before the booking changes. Check the underlying reservation and count the number of changes. When a service replies slowly, a second tool call must not create a second booking or charge.

For each run, log connected duration, audio sent and received, provider-reported billable usage where available, reasoning tokens, time to first audible reply, time to a useful answer and the final reservation state. Compare the median and a high percentile for delay across repeated calls. If the preview exposes no billable-audio counter, keep the cost forecast conditional and obtain the meter definition. A business deciding which calls to automate can use our [AI transformation services](https://www.digitalapplied.com/services/ai-transformation) to evaluate those workflows and their operating costs.

Save the record for the run where the caller changes the date while the lookup is pending. Listen to the response and inspect the tool result together. A pleasant conversation can still leave a customer with the wrong reservation, and a correct reservation can still take long enough that a customer would hang up.

### Base the switch on completed calls

MAI Realtime 1 is worth a Foundry trial for a team with an existing voice workflow, a measurable completion event and a reason to reconsider cost or responsiveness. Keep the current route available during evaluation. The signal to move traffic is a correctly completed call, at an acceptable delay and cost, under terms that fit the customer data involved.
