An AI phone agent costs about 4 to 15 cents per call minute at list prices we checked on 2026-10-13. At 3,000 minutes a month, that is roughly $160 to $450 a month for the running costs, depending mostly on which platform you pick, not which model.
The calculator below adds up the five things every call minute pays for: the platform, speech to text, the model, the voice, and the phone line. It runs in your browser and sends nothing anywhere.
Who this is for: owners and ops leads pricing an AI receptionist, booking line, or intake line before they talk to a vendor or a developer. Who should look elsewhere: if you need a human answering service, or if your calls are mostly long, open-ended sales conversations. The per-minute math still holds, but call quality will matter more than a few cents.
Model token assumptions
Every turn resends the instructions, tools, and the call so far, so input tokens grow through a call. These defaults are our assumption for a short booking call, not a measurement.
Monthly bill on Vapi
| Line | Per minute | Per month |
|---|---|---|
| Vapi platform | $0.0500 | $150.00 |
| Speech to text: Deepgram Nova-3 streaming | $0.0077 | $23.10 |
| Model: GPT-5.4 mini | $0.0078 | $23.42 |
| Voice: ElevenLabs Flash | $0.0125 | $37.44 |
| Phone line: Twilio, inbound to a US local number | $0.0085 | $25.50 |
| Plan, servers, and phone numbers | $1.15 | |
| Total | $0.0869 | $260.61 |
List prices, pay as you go, US calls. Voice prices checked 2026-10-13, model prices checked 2026-10-21. The total per minute includes the monthly fees spread over your minutes. Taxes, volume discounts, extra concurrency, and build or upkeep work are not included.
The same calculator is on its own page at /tools/voice-agent-cost-calculator/.
The answer for one booking line
These are the calculator’s starting values on all six options. Inbound US calls on a Twilio local number, Deepgram Nova-3 for speech to text, GPT-5.4 mini as the model, and ElevenLabs Flash as the voice wherever the platform lets you pick. Retell uses its own listed rates for GPT 5.4 mini and its platform voices. One phone number.
| Platform | Usage per minute | 1,000 min a month | 3,000 min a month | 10,000 min a month |
|---|---|---|---|---|
| Vapi | $0.086 | $88 | $261 | $866 |
| Retell AI | $0.103 | $104 | $309 | $1,026 |
| LiveKit Cloud | $0.036 | $88 | $161 | $466 |
| ElevenLabs Agents | $0.063 | $116 | $290 | $964 |
| Bland AI | $0.149 | $150 | $447 | $1,486 |
| Custom, self-hosted | $0.036 | $138 | $211 | $466 |
Prices checked 2026-10-13, from each vendor’s own pricing page (linked on the name). “Usage per minute” is before monthly fees. The monthly totals include plan fees: LiveKit Cloud on Ship ($50 a month, 5,000 agent minutes included), ElevenLabs Agents on Pro ($99 a month, 1,238 minutes included), and $100 a month for servers on the self-hosted line. That $100 is our placeholder, not a quote.
Two checks that the math is sane:
- Vapi’s own pricing page estimates $82 to $129 a month for 1,000 minutes with Deepgram, OpenAI, and ElevenLabs (Vapi, vendor estimate). Our default comes to $88, inside that range.
- Retell’s calculator shows “$0.11” a minute for its default setup (Retell). Ours is $0.103 because we swap in a cheaper model and a Twilio line.
Where the money goes
The parts under a voice agent are cheap. In the default scenario, speech to text is $0.0077 a minute, the model about $0.008, the voice about $0.0125, and the Twilio line $0.0085. That is under 4 cents a minute together.
The platform fee is often bigger than all of them. Vapi charges “$0.05/min Vapi hosting” (Vapi). Retell’s voice infrastructure is “$0.055/minute” (Retell). LiveKit Cloud charges “$0.0100/min” per agent session after the included minutes (LiveKit). Bland charges “$0.14/min” on its Start plan, which includes speech, model, and voice (Bland).
So at low volume, pick on ease of building and call quality, because the gap is a few dollars. At 10,000 minutes a month the gap between platforms is several hundred dollars a month, and it keeps growing with volume.
The assumptions you should change
Three inputs are our assumptions, not measurements. Open the calculator and set them for your calls.
- Model tokens per minute. We assume 6 model turns a minute, 2,500 input tokens and 60 output tokens per turn, and half of the input served from the prompt cache. Every turn resends the instructions, the tool list, and the call so far, so long calls cost more per minute than short ones.
- How much the agent talks. Voices are billed by characters. Cartesia says “One minute of audio generation requires 750-800 credits”, with 1 credit per character (Cartesia). We use 780 characters per spoken minute and assume the agent talks 40% of the call.
- Speech to text billing. AssemblyAI bills streaming on “the time the WebSocket connection is open, not the duration of audio sent” (AssemblyAI). Deepgram’s Nova-3 streaming list price is $0.0077 a minute, with a “Limited-time promotional” rate of $0.0048 (Deepgram). We use the list price.
What the calculator leaves out
- Building the agent. Prompts, tools that read your calendar or CRM, testing with real callers, and a handoff to a person. That is usually the biggest cost in the first year. A booking agent should also wait for human approval before it does anything that costs money. See how we price fixed-scope builds.
- Concurrency. Vapi includes 4 concurrent calls without a plan and charges “$10 / line / month” for more. Retell includes “20 Free Concurrent Calls”, then “$8.00/Concurrency/month” (Vapi, Retell).
- Speech-to-speech models. Retell lists GPT Realtime at “$0.345/minute” and GPT Realtime mini at “$0.07/minute”, replacing the model and voice lines (Retell).
- Taxes, volume discounts, recording storage, and compliance add-ons such as Vapi’s HIPAA add-on at “$2,000 / month”.
FAQ
How much does an AI receptionist cost per month?
For running costs alone, about $90 to $150 for 1,000 call minutes a month, and $160 to $450 for 3,000 minutes, at list prices checked 2026-10-13. Building and testing it is separate.
Is Vapi or Retell cheaper?
In our default scenario Vapi is cheaper per minute ($0.086 vs $0.103), mostly because Retell’s per-minute model rate is higher than paying OpenAI’s token price directly. With Retell’s cheapest listed model (GPT 5 mini at $0.008 a minute) the gap closes. The calculator lets you try both.
Is it cheaper to build my own voice agent?
The running cost is lower. With no platform fee, the default scenario is about $0.036 a minute plus servers. But you pay for building it and for someone to keep the servers and phone setup running. It starts to pay off at high, steady volume.
What is the cheapest phone line for an AI agent?
Of the carriers in the calculator, a Telnyx SIP trunk is cheapest: “Starting at $0.0032 per minute” inbound local (Telnyx). Twilio’s US local inbound rate is “$0.0085 / min” (Twilio). Web or app calls need no phone line.
Why does my bill not match the calculator?
Check the model token settings first. Long calls, big tool lists, and no prompt caching can raise the model line several times. Then check billing rounding. Retell tracks calls “to the nearest second”. Other vendors’ pages don’t say.
For what each part of a voice agent does, read what a realtime voice agent consists of. For a vendor-by-vendor look, read Vapi vs Retell vs LiveKit: which voice platform is cheapest. If the model is the part you want to tune, our AI cost per user calculator works through token math for chat and agent features, and should this feature be AI? helps decide whether a phone line needs a model at all. Dawloom builds AI agents with approval steps and a handoff to a person, from a written scope, billed hourly or by milestone.