—per minute
—per month
—from agent audio
Monthly cost by call volume
Audio scales linearly with minutes, so the monthly bill tracks call volume tightly. Cutting agent talk-time (the priciest tokens) or capping call length is the fastest lever.
| Calls / month | Audio minutes | Monthly cost |
|---|
Audio is the expensive token
Realtime voice models feel like magic — you talk, the model talks back, no transcription step you can see. But under the hood every second of audio in both directions is converted into audio tokens, and those tokens are priced well above text. A minute of speech is roughly a thousand to sixteen hundred audio tokens, and the audio the agent speaks is usually the single most expensive line item on the invoice. That is why an eight-minute voice call can cost more than an entire day of text chat: it is not the reasoning that is dear, it is the sound.
The practical takeaways fall out of the maths. Shorter agent responses save more than shorter user turns, because output audio is priced highest — a terse assistant is a cheap assistant. Capping maximum call length protects you from the long-tail calls that blow the budget. And if your interaction is transactional rather than conversational, a cascaded pipeline of speech-to-text, a text model and text-to-speech will almost always be cheaper, trading a little latency for a lot of savings. Model the cascaded alternative with the voice agent cost calculator and the speech-to-text calculator, and check the return against staff with the voice agent ROI calculator. Prices here are editable reference estimates — always confirm the current audio-token rates in your provider's pricing page.
Frequently asked questions
Why is the realtime audio API so much more expensive than text?
Because audio is billed as its own kind of token, and audio tokens are priced several times higher than text tokens. A speech-to-speech model like GPT Realtime or Gemini Live turns roughly one minute of speech into on the order of a thousand audio tokens, and both the audio you send in and the audio the model speaks back are billed — the output audio usually at the highest rate of all. A ten-minute voice call can therefore cost more than thousands of text chat messages. This calculator separates audio input, audio output and any text instruction tokens so you can see exactly where the money goes.
Is a realtime speech-to-speech model cheaper than STT + LLM + TTS?
It depends on volume and latency needs. A cascaded pipeline — a speech-to-text model, then a text LLM, then a text-to-speech model — usually costs less per minute because text tokens are cheap and transcription and synthesis are commoditised. A native realtime model costs more per minute but has much lower latency and keeps tone, interruptions and timing, which matters for natural conversation. The rule of thumb: use realtime where the conversation quality justifies the premium, and cascade where you are cost-sensitive or the interaction is more transactional. The calculator shows the realtime cost so you can compare it against a cascaded estimate from our voice-agent tools.
How many audio tokens is one minute of speech?
As a working figure, about one thousand to sixteen hundred audio tokens per minute of speech, depending on the model and how it encodes audio. The default here is a middle estimate you can override with the exact rate from your provider's docs. What matters most for a budget is that both directions are billed: the minutes the user talks are input audio tokens, and the minutes the agent talks are output audio tokens, which are typically the priciest line on the bill. Trim agent verbosity and you cut the largest cost.