What does AI voice actually cost?

Voice pricing is published in two different units — per million tokens for some models, per minute for others — which makes it almost impossible to compare at a glance.

Where a vendor publishes a per-minute rate, here it is, converted to the unit you actually budget in. All figures read from OpenAI's and Google's pricing pages on 17 July 2026.

Per minute, and per hour

The headline: transcribing an hour of audio costs about 18 cents. Doing it live costs about six times more, and translating live about eleven times more. Batch is cheap; realtime is the premium.

Audio tokens are expensive

For the conversational models billed per token, the same model charges wildly different rates by modality. Both from OpenAI's realtime line, per 1M tokens:

Audio input costs 8× text input on the flagship ($32 vs $4), and audio output about 2.7× text output ($64 vs $24). On the mini model the audio-to-text input ratio is worse still — nearly 17×.

This is the thing to internalise before building a voice product: the audio is not a delivery detail wrapped around a cheap text model. The audio is the bill.

The discount nobody mentions

Look at the cached column again. Audio input on gpt-realtime-2.1 drops from $32.00 to $0.40 — that is a 98.75% discount, or 80× cheaper.

For comparison, text caching on the same model runs $4.00 → $0.40, a 90% discount. Both are large. Audio caching is in a different category entirely.

We have not found this stated anywhere as a design principle, but the implication is clear enough: if your voice application replays the same audio context — a system preamble, a fixed greeting, a repeated prompt — caching it is the single highest-leverage cost decision available to you, by a wide margin.

The mini models are a real discount

gpt-realtime-2.1-mini is roughly a third the price of the flagship on audio ($10 vs $32 input, $20 vs $64 output) and six to ten times cheaper on text ($0.60 vs $4.00 input, $2.40 vs $24.00 output). That is a bigger spread than the equivalent gap in the text-model line.

As always, whether it is good enough is a question about your task, not about the table. But the spread is large enough to make trying it worth an afternoon.

What to actually do

1. Do not use realtime for a job that is not realtime. Transcribing recordings with a realtime model costs about 6× the batch rate for no benefit. Pick the endpoint that matches the job.
2. Cache your audio context. ~99% off is not a rounding error.
3. Budget on the audio rate, not the text rate. The text price on a voice model is the least relevant number on its row.
4. Try the mini model first. The spread is wide enough that the burden of proof sits with the expensive one.
5. Watch the units. A model quoted per minute and one quoted per million tokens cannot be compared without knowing your audio length. Convert before deciding.