How are realtime voice APIs priced?

By duration of audio in and out, typically per minute, rather than by text tokens. Speech-to-text, the model call, text-to-speech and an open session can each be billed separately.

Why do voice cost estimates come in low?

They model speaking time only. Pauses, latency, audio the caller interrupted, full-exchange retries and abandoned calls are all billed and usually left out of the model.

What is the biggest saving in a voice stack?

An aggressive idle timeout. Sessions held open while a user thinks or steps away are billed at full rate for producing nothing, and closing them is usually the largest single lever.

Quick answer: Every cost model built for text LLMs breaks on voice. Text is billed by tokens, and tokens are proportional to information. Audio is billed by duration, and duration is proportional to time —...

The unit economics of realtime and audio APIs

Updated September 1, 2026 · first published September 1, 2026

Every cost model built for text LLMs breaks on voice. Text is billed by tokens, and tokens are proportional to information. Audio is billed by duration, and duration is proportional to time — including all the time in which nothing useful is happening.

What you are actually paying for

A voice interaction is usually three or four billed things stacked. Speech-to-text on the inbound audio. The model call itself, which on realtime speech-to-speech APIs is priced per minute of audio in and out rather than per token. Text-to-speech on the way back, if the stack is not end-to-end. And often a separate charge for the session staying open.

That last one is the trap. A session held open while a user thinks, reads something on screen, or steps away is still a session. On a per-minute meter, dead air is a full-price product.

Where the estimates go wrong

Teams model voice cost as speaking time × rate and land well under actual. Four things are missing. Latency and pauses are billed but not modelled. Interruptions mean audio generated and paid for that the caller never heard. Retries and re-prompts replay the whole exchange, and a voice retry is far more expensive than a text one because it repeats the audio, not just the tokens. And abandoned calls cost the full duration to the point of hangup while producing nothing.

The metric that works

Cost per completed conversation, not cost per minute. Per-minute is the meter, not the unit of business value — and a change that shortens calls by improving the model may increase the per-minute rate while lowering the cost of getting a caller to an answer.

Then instrument the parts nobody sees: median and p95 session duration, ratio of billed to spoken time, abandonment rate, and cost of abandoned sessions as a share of the total. Close idle sessions aggressively; an idle timeout is usually the single largest saving available in a voice stack. And treat barge-in support as a cost feature as much as a UX one, since every second of speech a user talks over is a second you paid to generate.

Want this applied to your own LLM spend? FinOps LLM runs a free audit of your AI costs and shows where the savings are. Book free audit →

Back to research