PARAS MADAN
All writing

Latest in AI

MAI Realtime 1 Pricing: Why One Cent per Minute Is Not the Whole Bill

Microsoft's voice model adds separately billed reasoning to audio-minute pricing. Here is how to budget a pilot and distinguish preview access from production readiness.

Microsoft's MAI Realtime 1 announcement leads with a price that is easy to remember: $0.01 per audio minute. The important detail is that reasoning is billed separately. A voice-agent budget built only from call duration will therefore miss part of the model bill.

For a team considering a pilot, I would separate three decisions: whether you can access the model, whether it handles your conversations reliably and whether its complete cost works at your expected volume. The headline answers only part of the third question.

What the announced price includes

The launch announcement lists base speech-to-speech usage at $0.01 per minute, plus $1 per million reasoning input tokens and $5 per million reasoning output tokens. Microsoft describes OpenAI Realtime API compatibility and capabilities including interruption handling, tool use and prompted personas.

A useful budgeting expression is:

Model cost = audio minutes × $0.01 + reasoning input tokens ÷ 1,000,000 × $1 + reasoning output tokens ÷ 1,000,000 × $5

For example, 10,000 audio minutes produce a $100 base charge. If that hypothetical workload also uses 10 million reasoning input tokens and 2 million reasoning output tokens, the model total becomes $120. These are illustrative usage assumptions, not measurements or a forecast of typical consumption.

Preview access is not general availability

Microsoft offers a Foundry limited-access preview and says preview customers will not be billed until public preview begins. That is different from a publicly available production service at a permanently free price.

The model page also offers a personal demonstration through MAI Labs. Its terms displayed on the page restrict Labs to personal, noncommercial use, excluding school and work. A business evaluation should use the appropriate enterprise access route rather than treating the consumer demo as a work environment.

Before scheduling a customer-facing launch, confirm your own access, deployment options and service limits. A registration form is not an approved deployment, and an unbilled trial does not establish what production capacity will be available to you.

Measure reasoning rather than guessing it

Two calls of the same length can require different amounts of work. One might answer a simple opening-hours question. Another might compare options, inspect records and recover from a failed tool call. That is why I would record audio usage and reasoning usage separately during a pilot.

Group the results by task rather than averaging every conversation together. Track completed requests, abandoned conversations and human handoffs. Then calculate cost per successfully resolved request alongside cost per minute. A cheaper minute is not necessarily a cheaper outcome if people spend longer repeating themselves.

Keep retries and failed attempts in the dataset. Removing them makes the experiment look cleaner while hiding the behavior that will matter most in production.

API compatibility needs a regression test

Compatibility is a useful starting point, but it is not evidence that an existing application can switch models without changes. Test the actual events and behaviors your application relies on: interruption, cancellation, tool responses, session recovery and the transition to a human operator.

Use a recorded set of scenarios with permission from the participants, or scripted conversations created for testing. Include ambiguous requests, background noise and a caller changing their mind. Check whether an interrupted action is merely stopped in speech or also stopped in the connected system.

For consequential actions, keep confirmation and authorization inside the application. A natural-sounding response should not be the only evidence that the correct action occurred.

Keep the rest of the operating bill visible

The formula above covers announced model charges, not an entire voice service. Your application may separately incur telephony, infrastructure, storage, monitoring or human-support costs. List those categories even if some are initially zero.

Also budget engineering time for failure handling. A pilot that demonstrates a pleasant conversation is easier than a system that recovers safely when a service is unavailable. Make recovery part of acceptance testing instead of leaving it for after launch.

Make the rollout decision from your own calls

Microsoft's benchmark comparisons are reasons to investigate, not substitutes for workload evidence. Choose a bounded use case, define success before the trial and compare against your current workflow using the same scenarios.

I would expand only after the team can explain its resolution rate, handoff behavior and full cost. That turns an attractive starting price into a defensible operating decision.

I'm Paras. I work at the intersection of AI, engineering and distribution.

I help AI teams turn technical launches into practical stories developers can use.

Book a 30-minute call.