Every AI voice call is five services billed at once. Here's the per-layer math, a small calculator, and when owning the stack on LiveKit pays off.
The price on a voice AI platform's homepage is rarely the price on the invoice. That's because every call is five services billing at the same time, and platforms usually advertise only one of them.
I lead development on three live voice AI platforms (VoiceCake, Rawk.ai, and Dynaris), and this is the math I walk through whenever someone asks what an agent will cost to run. Rates below are US list prices I checked in September 2026. Providers change them often, so verify before you quote anyone.
The five layers of a voice call
| Layer | What it does | Example provider | Cost per minute |
|---|---|---|---|
| Telephony | Connects the agent to a phone number | Twilio | ~$0.0085 inbound, ~$0.014 outbound |
| Speech-to-text | Transcribes the caller in real time | Deepgram Nova-3 | ~$0.0077 streaming (list; $0.0048 promo) |
| LLM | Decides what to say and which tool to call | OpenAI, Anthropic, Google | ~$0.003 to $0.16, by model |
| Text-to-speech | Speaks the reply | Cartesia, ElevenLabs | ~$0.015 standard, ~$0.04 ElevenLabs |
| Orchestration | Runs the loop: turn-taking, interruptions, tool calls | Vapi, Retell, LiveKit Cloud | $0.05 Vapi, $0.055 Retell, $0.01 LiveKit |
Two layers dominate the bill: the voice and the model. A premium voice costs nearly three times a standard one, and a frontier model can cost many times a small fast one. For appointment booking, a fast mid-size model is usually the right call. Callers notice a slow reply long before they notice a slightly less clever sentence.
The orchestration layer is where platform choice shows up. Vapi and Retell charge about five cents a minute to run the conversation loop. LiveKit Cloud charges about one cent per agent-session minute once you're past your plan's included minutes, because you're running the agent code yourself.
A calculator you can run
Plug in your own rates and call volume:
# Per-minute rates, US list prices checked September 2026. Verify before use.
LAYERS = {
"telephony": 0.0085, # Twilio US inbound
"stt": 0.0077, # Deepgram Nova-3 streaming, regular list price
"llm": 0.03, # fast mid-size model (illustrative; varies widely)
"tts": 0.015, # standard voice
}
ORCHESTRATION = {
"vapi": 0.05,
"retell": 0.055,
"livekit_cloud": 0.01, # after the plan's included agent minutes
}
def per_minute(platform: str) -> float:
return sum(LAYERS.values()) + ORCHESTRATION[platform]
def monthly(platform: str, minutes: int) -> float:
return per_minute(platform) * minutes
def payback_months(extra_build_cost: float, minutes: int,
managed: str = "retell", owned: str = "livekit_cloud") -> float:
"""Months for orchestration savings alone to cover the extra cost of owning the stack."""
saving = (ORCHESTRATION[managed] - ORCHESTRATION[owned]) * minutes
return extra_build_cost / saving if saving > 0 else float("inf")
for name in ORCHESTRATION:
print(f"{name:<14} ${per_minute(name):.4f}/min")
# vapi $0.1112/min
# retell $0.1162/min
# livekit_cloud $0.0712/min
for minutes in (2_000, 10_000, 50_000):
print(f"{minutes:>6,} min/mo: Retell ${monthly('retell', minutes):,.0f} "
f"vs LiveKit ${monthly('livekit_cloud', minutes):,.0f}")
# 2,000 min/mo: Retell $232 vs LiveKit $142
# 10,000 min/mo: Retell $1,162 vs LiveKit $712
# 50,000 min/mo: Retell $5,810 vs LiveKit $3,560
print(f"{payback_months(5_000, 10_000):.1f} months") # 11.1 months
print(f"{payback_months(5_000, 2_000):.1f} months") # 55.6 months
A few things this simple model leaves out on purpose: LiveKit's plan fee and included minutes, Vapi's and Retell's concurrency limits, and phone number fees. Add them for your own case.
When Vapi or Retell is the cheaper choice
The payback numbers tell the story. At 2,000 minutes a month, orchestration savings take over four years to pay back a $5,000 difference in build cost. At that volume, a managed platform is the honest recommendation: you skip the engineering, launch in days, and the platform fee is small.
At 10,000 minutes, the same difference pays back in about 11 months, and at 50,000 minutes it's measured in weeks. That's where owning the stack starts to make financial sense.
Cost isn't the only reason to own it, though. On Rawk.ai, owning the pipeline is what let us cut end-to-end response time from about 900ms to about 320ms. You also keep control of call data, choose each provider independently, and don't have to rebuild if a platform changes its pricing.
Costs that show up on the first invoice
The per-minute math is never the whole bill. Budget 15 to 20 percent extra for the first couple of months because of:
- Test calls. Every test call bills like a real one. Tuning an agent before launch can burn hundreds of minutes.
- Calls that go nowhere. Voicemails, hang-ups, and spam still use telephony, speech-to-text, and orchestration.
- Transfers. A warm transfer keeps two call legs open, so telephony roughly doubles for those minutes.
- Monitoring. Recordings, transcripts, and observability are often billed separately, commonly around $0.01 per minute.
- Concurrency. Vapi includes 10 concurrent lines, then about $10 a month per extra line. Retell includes 20. A busy hour can hit the cap.
- Compliance. On Vapi, HIPAA is a published add-on at $2,000 a month on top of usage, which changes the math for clinical calls.
The buffer usually shrinks after launch, once testing stops and prompts settle.
The takeaway
- Under a few thousand minutes a month: use Vapi or Retell.
- Around 10,000 minutes and up: run the payback math for owning the stack on LiveKit.
- Either way: budget for the voice and the model first, because those two layers decide most of the bill.
The full version, including build-cost ranges and what moves them, is on keencraft.tech.
Top comments (0)