Microsoft's first speech-to-speech model costs $0.01 a minute

Microsoft AI unveiled MAI Realtime 1 during Ignite 2026. It is the company's first speech-to-speech model, built for real-time conversation in enterprise settings. Basic speech-to-speech usage starts at $0.01 per audio minute. When a task calls for deep reasoning, the model hands it off to a stronger reasoning engine, which is billed separately by token: $1 per million input tokens and $5 per million output tokens.

The model is compatible with OpenAI's Realtime API. In theory, an application already written against that interface can switch over by changing the endpoint.

Illustration: a wave links a headset and speech bubble, with a reasoning branch for Microsoft's real-time voice AI

Capabilities and rollout

The capabilities Microsoft lists lean toward customer service and call centers: users can interrupt the model, it drops in short acknowledgments such as “mm-hmm” and “okay” at the right moments (known in the industry as backchanneling), it holds up against background noise, it can call tools and enterprise systems to carry out actions, and its persona can be steered through prompts.

The rollout has several tracks. Developers can try the model in MAI Labs. Microsoft Foundry has opened a limited preview that requires an application; it is free during the preview, and billing starts only with the public preview. On the product side, the model is heading into Dynamics 365 Contact Center and Dragon Copilot, where the latter will use it to simulate patient conversations for nurse training. Microsoft named the travel platform Agoda and Reservations.ai as early users. The announcement page does not say which languages are supported, how many voices are available, or when the public preview begins.

The benchmark claim is a single sentence

Microsoft's own wording:

"Our model outperforms GPT Realtime 2.1 on quality, speed, and price"

The claim rests on Big Bench Audio (voice reasoning), τ-Voice (task completion for voice agents) and Microsoft's own Dynamics 365 customer-service evaluation. The announcement says MAI Realtime 1 sits on the “performance-price Pareto frontier” on the first two. It gives no actual scores, test configurations or latency figures; a footnote says only that results depend on specific evaluation configurations and load. Outsiders cannot currently verify the comparison.

Two ways to build a voice agent

The common architecture for voice agents has three stages: speech-to-text, then a text LLM, then speech synthesis. Chained in series, each stage adds latency, and a user cutting in mid-sentence is hard to handle. An end-to-end speech-to-speech model removes the intermediate transcription step. OpenAI's Realtime line takes this route, and Microsoft is now going head to head with it.

MAI Realtime 1's pricing separates two things: basic conversation is billed by the minute, deep reasoning by the token. An enterprise can first estimate the bulk of its cost from call duration and price the reasoning for complex questions separately, which is straightforward for a call center. The catch is that the $0.01 headline price is not the full cost of a call, and businesses that make many reasoning calls will need to measure it themselves.

Seen against Microsoft's wider lineup, it also released the coding model MAI Code 2 the same week. With voice now added, Microsoft has in-house models across coding, customer-service voice and foundation models; judging by the release cadence, the substitution element in its relationship with OpenAI is growing.

Sources: Microsoft AI official blog, CocoLoop, Ignite 2026 launch materials; checked against the per-audio-minute base price, the reasoning token rates and the first products to ship with the model.