What does an AI voice agent really cost per minute?
Rarely one number. A real per-minute cost stacks telephony, speech-to-text, the language model, text-to-speech and hosting, each billed by a different vendor.
The short answer
- A voice AI call's real cost is the sum of five components. Telephony (the phone line itself). Speech-to-text. The language model generating the response. Text-to-speech. And the hosting or orchestration layer running the pipeline.
- Each component is typically billed by a different vendor on a different unit. Per minute for telephony and speech services. Per token for the language model. Per character for text-to-speech. A single quoted rate can hide what is driving the total.
- The language model component usually has the widest possible range. Model choice, a fast, cheap model versus the largest available model, can change that one line item by several multiples.
- The honest way to evaluate a vendor's per-minute price is to ask what is included in it, not to compare the headline number alone. Two vendors quoting different numbers may be bundling different sets of components.
Why the headline number is not the whole picture
A voice AI call is not one system billed once. It is a pipeline. The call arrives over a telephony connection. The caller's speech gets converted to text. That text goes to a language model that decides what to say next. The response gets converted back to speech. Something has to host and orchestrate the whole sequence in real time, with low enough delay that the conversation feels natural. Every one of those five steps is typically a separate vendor or a separate line item, billed on its own unit.
A vendor's advertised per-minute price is either an all-in bundle of every component, or a quote for only their own piece of the pipeline with the rest assumed or passed through. Both are legitimate ways to price. They are not directly comparable without knowing which one you are looking at. That is the detail a headline number does not disclose on its own.
The five components of a voice AI call's cost
| Component | What it does | Typical billing unit |
|---|---|---|
| Telephony | Carries the actual phone call | Per minute |
| Speech-to-text (STT) | Converts the caller's speech into text the model can read | Per minute of audio |
| Language model (LLM) | Generates the agent's response from the conversation so far | Per token, input and output separately |
| Text-to-speech (TTS) | Converts the model's text response into spoken audio | Per character or per minute |
| Hosting/orchestration | Runs the real-time pipeline connecting the other four | Per minute or a flat platform fee |
Every vendor quote should be checked against which of these five it actually includes.
How to actually evaluate a quoted rate
Ask a vendor directly which of the five components their quoted rate includes. Get a dated source for each: the telephony carrier's own pricing page, the STT and TTS provider's published rate, and the language model provider's per-token price for the model used. A quote that cannot be broken down this way is harder to trust. Not because it is necessarily wrong, but because it cannot be compared against a competing quote with confidence.
Model choice deserves particular attention. It tends to be the widest-ranging single line item. A faster, lower-cost model can process a turn for a fraction of what the largest available model costs, and that difference multiplies across call volume. A vendor defaulting to the most capable model everywhere is not necessarily giving a better product. It may just be pricing higher than the conversation quality requires.
Disclosure: SalesCrew is our product, and outbound and inbound voice AI are on our roadmap and not shipped today. This page is not a quote for a SalesCrew rate. It is a breakdown of what any voice AI vendor's per-minute number is made of, useful for evaluating any provider.
Compare what is included, not only the headline rate
Questions
- Why do vendor pricing pages show one number?
- A single per-minute figure is easier to sell against, and some vendors bundle every component into one rate. The number is not necessarily wrong. It is only comparable to another vendor's number if both bundle the same components the same way, which is rarely disclosed.
- Does a longer call cost proportionally more?
- Mostly yes. Telephony, speech-to-text and text-to-speech scale with call duration. The one component that can scale non-linearly is the LLM, if the conversation's context grows long enough to change how many tokens each turn costs to process.
- What is the single biggest lever for lowering cost per minute?
- Model choice for the LLM component usually has the widest price range of any single line item. Faster and cheaper models cost much less per token than the largest, highest-quality ones, and the gap compounds across many calls.