Voice AI Cost Per Minute Calculator

What this tool does

A live voice agent is four services running at once. This calculator asks for the rate of each and shows you the combined per-minute figure, the split between layers, and what a month of calls comes to at your expected volume.

The four layers are telephony (the actual phone call), speech-to-text (transcribing the caller), the language model (deciding what to say) and text-to-speech (saying it). Platform fees sit on top. Because every provider prices differently and rates change, all of them are inputs — pull yours from the pricing pages you are actually on, such as Retell AI, Vapi, Twilio voice and the OpenAI API.

A worked example

You are quoted a per-minute platform rate and assume that is the cost. Then you add the inbound telephony leg, which is billed separately by your carrier. Then the transcription, which is usually per audio minute and runs the whole call length. Then the model, which is billed per token and depends on how long your system prompt is and how much conversation history you resend on every turn. Then the voice synthesis, billed per character of output.

The pattern this reveals is consistent: on a short, well-scoped call the platform and telephony dominate; on a long call with a large system prompt, the model layer overtakes everything, because you resend the whole context on every single turn. That is the number worth knowing before you write a 3,000-word prompt.

The four layers

Where people go wrong

Budgeting from the platform rate alone. It is the headline number precisely because it is the smallest part of a real bill for most builds.

Writing an enormous system prompt. The prompt is resent with every turn. A 3,000-token prompt on a twenty-turn call is 60,000 input tokens before the conversation itself is counted. Trimming it is usually the single largest cost reduction available. How to Write a System Prompt for AI Agents covers doing that without losing behaviour.

Forgetting that telephony rounds up. A hundred 15-second calls bill as a hundred minutes on the carrier leg.

Ignoring failed and abandoned calls. A caller who hangs up after four seconds still consumed a telephony minute and a model turn. At high volume these add up.

Assuming a cheaper model is always cheaper. A faster, smaller model that needs three turns to do what a larger one does in one can cost more overall, and it makes the call longer, which costs you on every other layer too.

FAQ

Which layer usually costs the most?

It depends on call length. Short, transactional calls are dominated by telephony and platform minutes. Long, conversational calls are dominated by model tokens, because context is resent on every turn.

Does the calculator include the carrier leg?

Yes, as its own input. Take the rate from your carrier's pricing page for the destination country — inbound and outbound differ.

How do I estimate token usage per minute?

Count your system prompt once, estimate turns per minute (two to four is typical), and multiply the prompt plus the accumulated history by the turn count. The calculator does the arithmetic once you supply those assumptions.

What about the cost of a human taking the call instead?

That is the comparison that matters commercially rather than technically. The AI receptionist cost calculator sets the two side by side.

Do these figures include VAT or sales tax?

No. Enter rates exclusive of tax and add it to the total, as every provider bills separately.

Next steps

If you are choosing a platform, Retell AI Review covers one of the main options in detail and How to Use Retell AI covers the build. For the carrier side of the bill, use the Twilio cost calculator.

Want the stack costed against your actual call volumes? Book a discovery call.