A missed call is the cheapest lead a small business ever loses. The caller wanted a quote, got voicemail, and dialled the next name on the list. AI phone agents exist to answer those calls, and in 2026 they cost less per minute than most owners assume. The problem is that nobody sells them by the minute in a way you can compare.
The short answer: a voice agent runs about $0.03 to $0.09 a minute on Vapi once you add the speech, model and phone line together, and $0.07 to $0.31 a minute on Retell. A three-minute call therefore costs between 12 and 93 cents. The platform is never the expensive part. The setup and the minutes you waste on calls it cannot close are.
Why per-minute pricing hides the real bill
Every voice agent is four products wearing one name: speech recognition, a language model, a voice, and a phone line. Vapi and Retell both pass those parts through at cost and charge you for the minutes the agent runs. So the headline rate depends entirely on which parts you pick.
This is the same trap that made automation pricing hard to compare, which we worked through in what each automation actually costs per task. The unit is not the subscription. The unit is the minute or the call, and the cheap-looking plan can be the expensive one at your volume.
Vapi: what the numbers on the page add up to
Vapi does not sell a bundle. It lists components, and on the day we read the pricing page the numbers were:
- Transcriber, Deepgram: $0.0095 to $0.0099 per minute
- Intelligence model, OpenAI: $0.0077 to $0.0452 per minute
- Voice, ElevenLabs: $0.0146 to $0.0238 per minute
- Phone line: Twilio at $0.008 per minute, or Telnyx at $0.00814 per minute
There is no platform fee on the usage plan. You start with $5 in free credits and pay only for minutes that run. If you want more support and concurrency, Core is $29 a month with 5 phone numbers, 10 concurrent calls and 30 days of raw data. Pro is $999 a month minimum and then 10% of the Vapi hosting fee, which only makes sense once you are running real volume.
What the three lines add up to
Add the cheapest sensible stack together and you get $0.0095 plus $0.0077 plus $0.0146, which is $0.0318 a minute. Add the phone line and it is $0.0398. Call it 4 cents a minute.
Retell: one range, four components inside it
Retell does the opposite. It quotes a single figure, $0.07 to $0.31 per minute for AI voice agents, with no platform fee, no minimum and $10 in free credits. The spread comes from the same four parts, priced inside the platform:
- Retell voices: $0.055 per minute
- OpenAI voices: $0.015 per minute
- ElevenLabs Multilingual v2: $0.040 per minute
- ElevenLabs v3 Conversational and v4 Turbo: $0.10 per minute
- LLM, standard fast tier: $0.015 per minute
Cheap voice plus cheap model lands you near the bottom of that range. A premium ElevenLabs voice with a heavier model lands you near the top. Neither number is wrong. They just describe different products.
The three-minute call, costed three ways
One call, four ways
Take the call every small business gets: someone asks about price, wants to know if you cover their area, and books a time.
- Vapi, cheapest stack: $0.12 per call
- Vapi, top of the published ranges: $0.26 per call
- Retell, $0.07 a minute: $0.21 per call
- Retell, $0.31 a minute: $0.93 per call
Twenty calls a day is 600 calls a month and about 1,800 minutes. On the cheapest Vapi stack that is roughly $72 a month, plus $29 if you take Core. On Retell it is $126 to $558 depending on the voice and model you choose. Same job. Very different bill.
What a human costs for the same three minutes
Do the honest version of this comparison, because vendors do not. A person is not paid per call. At $18 an hour for a part-time receptionist covering 20 hours a week, you are paying $1,560 a month before tax, and that person answers every call in those hours, including the four-minute ones and the ones that go nowhere.
Per answered call, if they handle 12 calls an hour, the wage share is $1.50. The AI agent at 12 to 93 cents is cheaper per call at every single tier.
What the person gives you that the agent does not is judgement on the twentieth odd question of the day, and that is what you are really buying.
The number that decides it: how many calls you miss
Count before you buy
Before any of this matters, count your missed calls for one week. Most owners guess, and the guess is low. If you miss 30 calls a week, an agent pays for itself the first time it books one job worth more than $72.
If you miss four calls a week, you are spending $72 to solve a $20 problem.
Retell and Vapi both charge for minutes, not calls, so a week of 30 three-minute calls is 90 minutes. That is under $4 on the cheapest stack. The monthly plan only starts to matter once you are past a few hundred calls.
Where the cost quietly grows
Three things push the bill up after month one, and none of them are on the pricing page.
- Long calls. Every extra minute of "let me just check with my manager" costs the same as a good minute. A transfer to a human still bills the AI time first.
- Interruptions and silence. A confused caller who repeats themselves doubles the length of the call, and the meter does not care why.
- Phone numbers. Vapi charges $10 per line per month beyond what your plan includes. Six locations means six lines.
The setup work nobody budgets for
The minutes are cheap. The build is not. A voice agent that sounds human needs a written script with branches, a list of questions it must never guess at, a transfer rule, and a calendar that actually accepts bookings.
Expect a day of work for a simple inbound receptionist and a week for anything with qualification logic.
If you skip the script, you get an agent that says "I can help with that" and then does not. That is worse than voicemail, because the caller believes a person will call back.
When an AI agent is cheaper than a person
- You miss calls after hours and on weekends, when nobody is paid to answer.
- Most calls are the same five questions, and the answers fit on one page.
- You need two languages on the same number without hiring twice.
- Your call volume swings hard and you cannot staff to the peak.
When a person is cheaper than an AI agent
- Your calls are consultations, not bookings. Judgement is the product.
- Your customers are older, or calling about something emotional, and a bot on the line costs you the job.
- You already have someone at a desk who is not fully loaded.
- You cannot maintain the script, so the agent drifts and starts inventing answers.
What the vendor pages do not say
Two gaps are worth naming. First, neither page publishes one final all-in per-call figure, because there is no single one. The range is the product. Second, the model component is where your bill actually moves: on Vapi the intelligence model alone ranges from $0.0077 to $0.
0452 a minute, which is a six times spread on the same call.
We could not read Synthflow's pricing table from a plain fetch, because the page renders prices client side. Treat any comparison that quotes it without saying so as second hand.
A five-step test before you pay for a month
- Export last month's call log and count missed calls and average minutes. Real numbers, not feelings.
- Write the five questions your callers actually ask, in their words.
- Build one agent, inbound only, with one transfer rule.
- Run it on the free credits for a week and listen to every recording, not a sample.
- Compare the booked-job value against the minute bill before you add the second use case.
If you only miss five calls a week
Do not buy a voice agent yet. Turn on a missed-call text back, add your hours to your Google profile, and put a booking link on the voicemail. That costs nothing and covers most of the gap.
Come back to the agent when the log shows forty missed calls and two of them asked for a quote.
The transfer trap
A transfer feels free because the AI stops talking. It is not. The agent bills the minutes before the handoff, the human bills from the moment they pick up, and the caller spends the gap listening to hold music wondering if the call dropped.
Every transfer you design should have a stated condition, like "price above $2,000" or "existing customer with an open job", and a named person on the other end during business hours.
What a good agent sounds like
Short sentences. It says the business name in the first three seconds. It asks one question at a time. It admits what it does not know and offers the transfer instead of guessing. It never repeats the caller's full sentence back to them.
If your test call sounds like a call centre reading a script, the problem is the script, not the model.
Verdict: At 12 to 93 cents for a three-minute call, the technology is not the expensive part of the decision. The expensive part is building one that does not embarrass you, and paying for it on a volume you do not have. Count the missed calls first. If the number is above thirty a week, a Vapi stack at 4 cents a minute will pay for itself in a month. If it is under ten, fix your voicemail instead.
Questions people ask
How much does an AI phone agent cost per call?
About 12 to 93 cents for a three-minute call. The low end is a cheap speech-to-text, model and voice with a Twilio line. The high end is a premium ElevenLabs voice on Retell.
Is Vapi or Retell cheaper?
Vapi publishes component prices that add up to roughly 4 cents a minute on the cheapest stack. Retell quotes a single all-in range of $0.07 to $0.31 a minute. Neither is wrong, they package the same four parts differently.
Is an AI receptionist cheaper than hiring someone?
Per call, yes. At $18 an hour and 12 calls an hour, a person carries about $1.50 of wage per call against 12 to 93 cents for the agent. But a person is paid for the shift, not the call, so the comparison only holds if you actually have calls to answer.
Do I need a separate phone number?
Yes. Both platforms connect through a phone line, and Vapi charges $10 per extra line per month beyond the numbers included in your plan.
What if I only miss a few calls a week?
Do not buy an agent yet. Turn on a missed-call text back, put your hours on your Google profile and add a booking link to the voicemail. Revisit when the log shows around 40 missed calls a month.
Related: what each automation actually costs per task, and the AI cost calculator for your current stack.