The Real Cost of an AI Voice Agent, Per Minute

Every voice AI pricing page quotes a platform fee, not a price. Here is what a voice minute is actually made of, how to work out any vendor’s real rate — and ours, published in full, including the parts that do not flatter us.

Every voice AI pricing page in this market quotes a number that is not the price. Vapi says $0.05 a minute. That is the platform fee — what Vapi charges for orchestrating a call. It does not include the language model that decides what to say, the speech-to-text that hears the caller, the text-to-speech that produces the voice, or the telephony that carries the audio. Those four are usually more than the platform fee combined.

This is not a hit piece, and Vapi is a capable platform. The problem is structural and the whole category shares it. This article is the arithmetic: what a voice minute is actually made of, how to work out any vendor’s real rate, and — since we are asking you to trust our numbers — ours, published in full, including the ones that do not flatter us.

The headline number is never the number

There are three pricing models in this market, and knowing which one you are looking at tells you immediately how much arithmetic you still have to do.

The trap is comparing across models. A $0.05 platform fee against a $0.25 bundled rate looks like a five-fold difference and is frequently no difference at all. You have to reduce everything to one number — total spend divided by minutes talked — before any comparison means anything.

What a voice minute is made of

Five things happen on every minute of an AI phone call, and somebody bills for each of them.

Why long calls cost more than twice as much

This is the part that surprises people on component pricing. On turn twenty, the model is not just reading the last thing said — it reads your system prompt and the entire conversation so far, then produces a reply. Cost per turn rises as the call goes on.

Two consequences follow. A 900-word system prompt is not free; it is a tax on every turn of every call, in latency as well as money. And a booking flow that takes eight turns instead of four does not cost twice as much — it costs rather more.

Working out Vapi’s real rate

Take a common configuration: GPT-4o for reasoning, Deepgram Nova-2 for transcription, ElevenLabs Turbo v2 for the voice, Twilio for telephony.

So the real figure is three to five times the advertised one, and it arrives as five line items on separate invoices with separate billing cycles. At 10,000 minutes a month that is $1,400–$2,500 rather than the $500 the pricing page implies.

The spread is the actual problem

Notice the range: $0.14 to $0.25 is a 79% spread on the same configuration. Where you land depends on how long your calls run, how much the agent talks, and how long your prompt is. That is not a number you can put in a budget, and it is not a number you can quote a client.

Retell, Bland and rolling your own

The build-it-yourself maths

Assembling the stack yourself removes every platform fee, which looks compelling on a spreadsheet. What the spreadsheet omits is that the platform fee is not buying you an API wrapper. It is buying turn-taking that knows when a caller has finished speaking, interruption handling, voicemail detection, retry logic, call recording, transfer plumbing and the telephony integration — plus somebody to be awake when it breaks at 2am.

A rough rule: below about 100,000 minutes a month, the engineering time to build and maintain that costs more than the margin you save. Above it, and with a team that already knows telephony, the calculation changes. If you want the components without the assembly, the Vociply API sits between the two.

What Vociply charges, in full

We publish these because asking you to distrust someone else’s pricing page while being vague about our own would be worth nothing. These are the real figures from our price book, including the ones that do not flatter us.

United States

Kenya and Nigeria

These markets are priced independently rather than converted from the dollar figure. The monthly commitment is deliberately small so a business can start without a large outlay, with a proportionally smaller bundle.

Full detail on how minutes are counted, what happens when you exhaust a bundle, and how the destination affects the overage rate is in the billing documentation, and current pricing for your market is on the pricing page.

Where stacked pricing actually wins

Component pricing is not a trick. There are cases where it is the correct choice, and pretending otherwise would make everything above less credible.

• You have negotiated provider rates. If your company already has an enterprise agreement with OpenAI or ElevenLabs, bringing those keys removes a markup nobody else can remove for you.
• You want per-agent model selection. A cheap fast model for the FAQ line and something heavier for complex sales, tuned independently.
• You are experimenting rather than operating. Swapping voice providers weekly to find the right one is exactly what component pricing is good at.
• You are very large. Above roughly 100,000 minutes a month, a percentage point of margin is worth managing five vendor relationships for.

The counter-argument is that every one of those is an ongoing operational commitment, and that most teams doing 5,000 minutes a month would rather spend the time on what the agent says than on which vendor changed their per-character rate.

Calculating your own number

Do this before you talk to any vendor, including us. It takes fifteen minutes and it makes every subsequent conversation shorter.

calls per day
  × average call length in minutes
  × working days per month
  = talk minutes

then add:
  + retries on unanswered outbound calls (typically 20-40% more attempts)
  + voicemails reached before detection stops the call
  + time on calls that were transferred to a human, which keep running
  + inbound calls you did not model at all

The three additions people forget are worth 20–50% between them. A campaign modelled at 2,000 minutes routinely bills 2,600.

Then ask every vendor the same five questions

• What is the all-in cost per minute, including model, speech and platform?
• What is billed separately, and on whose invoice does it arrive?
• How does the rate change on a ten-minute call versus a two-minute one?
• What does a call to my most expensive destination cost?
• What happens when I exceed my plan — does it stop, or does it bill?

A vendor who cannot answer the first question with one number is quoting you a platform fee. That is not disqualifying — but you now know how much arithmetic is still yours.

The costs nobody budgets for

Kenya, Nigeria and the parity trap

Most voice AI vendors price in dollars and convert. That produces an entry point of several hundred dollars a month, which in Nairobi or Lagos is not a pilot budget — it is a hiring decision. The result is a category that is effectively unavailable to the businesses with the most phone traffic per employee.

The alternative is to price each market on what it can bear and shrink the bundle to match, rather than discounting the rate below cost. That is why Kenyan Essentials is KSh 2,600 for 75 minutes rather than a converted $299 for 1,200: a low commitment to start, at an effective rate that still covers what the call costs to deliver.

Billing is in your market’s own currency, and the overage rate is too — a Kenyan account calling a Nigerian mobile is priced in shillings against a shilling plan rather than converted through the dollar.

What real deployments spend

Rates are abstract. These are the shapes of actual usage, and they are more useful for budgeting than any per-minute figure.

Two of these are documented end to end if you want to see the mechanics: instant lead callback and an AI receptionist that books appointments.

Before you sign anything

• I have one all-in per-minute figure from every vendor on my shortlist.
• I know which costs arrive on someone else’s invoice.
• I have added retries, voicemails and transfer time to my minute estimate.
• I know what happens when I exceed the plan — stop, or bill.
• I have checked the rate for the countries I actually call.
• I have counted engineering time in any build-versus-buy comparison.
• I have run a pilot and measured real call length rather than assuming it.

If you want the feature-by-feature version of this comparison rather than the pricing one, we maintain Vociply versus Vapi. And if per-minute rate is the whole decision at serious volume, talk to us about Enterprise — dedicated infrastructure changes the arithmetic above.