What an AI phone agent actually costs per minute

The short answer

Managed platforms quote seven to thirty one cents a minute, and most small shops land near the low end.

Retell's pricing page states an all in range of $0.07 to $0.31 per minute for its voice agent, counting its platform fee, the language model and add ons. Vapi's pricing calculator works an example at 1,000 call minutes a month and lands at $82 to $129 before phone carrier charges, which is about $0.08 to $0.13 a minute. Deepgram's price list puts its all in Voice Agent API at $0.075 a minute on the standard tier, not counting the phone line. So a shop that answers 1,000 minutes a month is looking at roughly $100 to $150, going by those three vendor pages.

Where the money goes

Every call is a stack of small charges, and this table shows each one.

Every figure above comes from the named vendor's public price page, and the two marked as my estimate are my own arithmetic on those vendor rates, not the vendor's claim.

The platform fee is the big part

Most of the per minute price is hosting, not the AI itself.

Add up the raw parts. Twilio's inbound line at $0.0085, Twilio's Media Streams at $0.0044, Deepgram's Nova-3 hearing at $0.0048, a cheap OpenAI model at about $0.001 by my token estimate, and Deepgram's Aura-2 voice at about $0.012 by my character estimate come to under 3 cents a minute, all from those vendors' price lists. Then Vapi's pricing page adds $0.05 a minute just to host it, and Retell's page adds $0.055. The hosting fee alone is close to double the raw parts. You are paying someone to wire it together and keep it running, which is fair, but know that is what you are buying.

Model choice swings the price

The same platform can cost 200 times more per minute depending on the brain you pick.

Retell's pricing page lists its cheapest model, GPT-5 nano, at $0.0016 a minute and its priciest, the GPT-5.5 fast tier, at $0.32 a minute. That is a 200 times spread on one page. Middle picks on the same Retell page, GPT-5.6 Terra and Claude 5 Sonnet, both run $0.064 a minute. Vapi's page reports its customers' OpenAI cost at $0.0077 to $0.0452 a minute. Anyone who quotes you one number is quoting one setup, not a market price.

Voice quality costs extra

The voice that sounds human is the most expensive line on the bill.

Deepgram's price list sells Aura-2 text to speech at $0.030 per 1,000 characters. At about 150 words a minute of talk, with the agent talking half the call, my rough math comes to about $0.012 per call minute, and that math is mine, not Deepgram's. ElevenLabs sells by monthly plan instead, and its own pricing table puts its voices at about $0.17 to $0.36 per minute of generated speech. Vapi's page says its customers actually pay ElevenLabs $0.0146 to $0.0238 a minute through Vapi. Either way, the cheap voice sounds cheap, and the good voice can cost more than the whole rest of the call.

Fixed costs nobody mentions

Per minute math hides the monthly bills and the setup work.

A phone number rents for $1.15 a month on Twilio's price list and $2.00 a month on Retell's. ElevenLabs' plans start at $11 a month for Creator, then $99 for Pro and $299 for Scale, by its pricing page. None of these vendor rates include writing the prompt, testing it, wiring it to your calendar or your job software, or tuning it each month. For a shop doing a few hundred minutes a month, that setup and upkeep costs more than the minutes do.

Compared to a receptionist

A person costs about eighteen dollars an hour before taxes, but the two are not the same job.

The Bureau of Labor Statistics wage survey for receptionists and information clerks puts the median pay near eighteen dollars an hour, and I have not rechecked that against the newest table, so treat it as a rough figure. Employer payroll taxes and benefits add roughly 25 to 30 percent on top, by my own loaded cost rule of thumb, not by any survey. At 1,000 answered minutes a month, the agent costs about $100 to $150 by the Vapi and Retell ranges above, which is less than one week of pay at that BLS rate.

But the receptionist is paid for the whole shift, including the quiet hours. The agent bills only for talk time. That makes the agent look cheaper than it is when calls are few and the setup bill is large. Human answering services sell in monthly minute bundles, and I did not open a rate card, so I will not print a per minute number for them.

What stays human

Someone still has to handle walk ins, paperwork and the calls that go sideways.

An agent can answer, take a message, book a slot and pass a call on. It cannot greet the person who walks in the door. It cannot sign for a delivery or sort the mail. It cannot decide that an angry customer needs the owner right now unless you wrote that rule ahead of time, and it will not catch the rules you forgot to write. Plan on a person owning the escalations and reading the call log at least once a week.

What we got wrong

I started out thinking this was a price question, and it is a setup question.

I expected to find one market rate. Instead I found that the per minute figure is mostly the platform fee, and the AI parts under it add up to under 3 cents by the vendor lists above. I also found that list prices understate what you pay. Vapi's page shows its customers paying Deepgram $0.0095 to $0.0099 a minute, about double Deepgram's own Nova-3 list price of $0.0048. And every price here came from a vendor page that can change any week, so check the source links before you sign anything.

If you would rather not do this math yourself, have us install the phone agent for you. Book the $199 consult and we will price your real call volume against the table above before anything is bought.