Microsoft's First Speech-to-Speech Model Costs One Cent a Minute, With Reasoning Billed Separately
Microsoft introduced MAI Realtime 1 at Ignite 2026, its first model that inputs and outputs speech directly. The advertised audio price is one US cent per minute. Tasks that call a separate reasoning model incur additional token charges, so the audio rate is only one part of the bill.
How the interaction works
A traditional voice assistant chains transcription, a language model, and speech synthesis. Passing audio directly can help preserve cues such as when to wait, acknowledge, or yield to an interruption. Microsoft also describes routing harder requests to a reasoning engine while trying to keep the dialogue responsive.
What 57.5% means
The source article reports a 57.5% task-completion score in the evaluation discussed by Microsoft. This describes performance on that benchmark rather than the likelihood of success across all customer calls. A fluent conversation and a correctly completed transaction should be measured separately.
How to compare costs
Measure reasoning usage, repetitions, transfers to staff, and completion rates with realistic calls. Cost per correctly completed task is more informative than the advertised audio price for workflows involving billing, permissions, or complex decisions.