Clients who ask for an "AI receptionist" usually ask the same second question: what will it cost us each month? The honest answer is "it depends on minutes", which is true and useless. So we started giving them a number from a tiny script instead.
This post walks through how voice AI platforms bill, which parts of a call you actually pay for, and a 60-line Python estimator you can adapt for your own quotes. Prices below are the vendors' published rates, checked on 19 November 2026, in US dollars. They change often, so treat them as inputs, not constants.
How a voice agent call is billed
A phone call to a voice agent has four moving parts, and per-minute platforms bill most of them separately:
- Voice infrastructure: the platform that streams audio, handles turn-taking and interruptions.
- Speech: speech-to-text on the way in and text-to-speech (the voice) on the way out.
- The model: the LLM that decides what to say and which tool to call.
- Telephony: the phone number and the carrier minutes.
Retell AI publishes each part. A typical configuration on its price list: voice infrastructure $0.055/min, a standard voice $0.015/min, GPT 6.1 Terra $0.032/min, and telephony $0.015/min in the US and UK, plus $2/month per number. Retell's own headline range is $0.07–$0.31/min depending on the voice and model.
Vapi charges a $0.05/min platform fee and you pay the providers you plug in (speech, model, carrier). If you bring your own carrier, Twilio lists US local numbers at $1.15/month and inbound local minutes at $0.0085.
Three billing details catch people out:
- Silence and hold time are billed. A caller who says "hang on, let me find my booking number" costs you money.
- Very short calls have a floor. Retell bills dynamic-greeting calls under 10 seconds as 10 seconds.
- Long prompts can cost extra. Retell scales billed time up when the average prompt goes over 4,000 tokens, so a bloated system prompt isn't free.
The estimator
The script takes a call profile (calls per month, average minutes, unique callers) and compares a per-minute build with subscription pricing.
from dataclasses import dataclass
@dataclass
class PerMinuteStack:
name: str
infra: float # $/min platform or voice infrastructure
voice: float # $/min text-to-speech
model: float # $/min LLM
telephony: float # $/min carrier
numbers: float # $/month for phone numbers
addons: float = 0.0 # $/min knowledge base, denoising, PII removal...
def per_minute(self) -> float:
return self.infra + self.voice + self.model + self.telephony + self.addons
def monthly(self, minutes: float) -> float:
return minutes * self.per_minute() + self.numbers
@dataclass
class CallProfile:
calls: int
avg_minutes: float
unique_callers: int
hold_buffer: float = 0.10 # extra billed time for silence and hold
@property
def billed_minutes(self) -> float:
return self.calls * self.avg_minutes * (1 + self.hold_buffer)
def per_caller_plan(profile, price, included, overage):
extra = max(0, profile.unique_callers - included)
return price + extra * overage
def per_call_plan(profile, price, included, overage):
extra = max(0, profile.calls - included)
return price + extra * overage
retell_typical = PerMinuteStack(
name="Retell, standard voice + GPT 6.1 Terra",
infra=0.055, voice=0.015, model=0.032, telephony=0.015, numbers=2.00,
)
profile = CallProfile(calls=300, avg_minutes=3, unique_callers=220)
print(f"Billed minutes: {profile.billed_minutes:.0f}")
print(f"{retell_typical.name}: ${retell_typical.monthly(profile.billed_minutes):,.2f}"
f" (${retell_typical.per_minute():.3f}/min)")
print(f"Per-caller plan ($129, 500 included): ${per_caller_plan(profile, 129, 500, 0.26):,.2f}")
print(f"Per-call plan ($500, 300 included): ${per_call_plan(profile, 500, 300, 2.17):,.2f}")
Run it with 300 calls of 3 minutes from 220 different callers and a 10% buffer for hold time:
Billed minutes: 990
Retell, standard voice + GPT 6.1 Terra: $117.83 ($0.117/min)
Per-caller plan ($129, 500 included): $129.00
Per-call plan ($500, 300 included): $500.00
The two plan shapes are real: Goodcall's Growth plan is $129 for 500 unique customers, and Smith.ai's AI receptionist has a $500 plan for 300 calls. Notice the units. A per-caller plan ignores repeat calls; a per-call plan doesn't. For a clinic where the same patients call several times a month, that difference matters more than the headline price.
What the per-minute number leaves out
The estimator covers running cost. It ignores the build, which is where most of the engineering goes:
- Tool calls. Booking into the client's calendar or practice-management system, looking up a job or order, writing a call summary to the CRM. Each integration needs auth, error handling and tests.
- Hand-off. Rules for when to transfer to a human, warm vs cold transfer, and what happens out of hours.
- Evaluation. Replaying real (consented, redacted) call recordings against the agent before launch, then sampling live calls each week.
- Disclosure and recording. Telling callers they're speaking to an AI where local rules require it, and handling call-recording consent per state or country.
Model choice is the biggest lever on running cost. Swap the model line for a premium voice and a larger model and the per-minute figure can double or triple. Run the numbers with two or three model options before you commit, and check whether the cheaper model still handles your hardest real calls.
Tips for keeping the bill sane
- Keep the system prompt short and move reference content into a knowledge base the agent queries.
- Put a hard time limit on calls and a polite wrap-up when it's reached.
- Detect voicemail and silence early on outbound calls.
- Cap concurrency to what the business actually needs; extra lines are a monthly fee.
- Log billed minutes per call and alert on outliers (a stuck call can run for an hour).
If you need help scoping a voice agent that books into your own systems, our team does AI voice agent development for businesses in the USA, UK, Canada and Australia.
FAQ
What does an AI voice agent cost per minute?
On Retell AI, published rates add up to $0.07–$0.31 per minute depending on the voice and model. On Vapi it's a $0.05/min platform fee plus the speech, model and carrier you choose.
Is a subscription AI receptionist cheaper than building one?
At low volume, often yes, because there's no build cost. A custom agent makes sense when it has to work with your own systems or you handle thousands of minutes a month.
Are hold and silence billed?
On per-minute platforms, yes: billing covers the whole connected call.
How accurate is this estimator?
It's as accurate as its inputs. Use your real call logs for call counts, durations and unique callers, and re-check vendor prices before quoting.
Originally published on the Dzimark blog.
Top comments (0)