LiveKit pricing for voice agents: the agent-session minute is one line of the bill
LiveKit Cloud charges $0.01 per agent-session minute after your plan's included minutes. Models, phone minutes and recordings are separate meters. Here are two monthly bills at the same volume, one for a web app and one for a phone line.
LiveKit Cloud has four plans: Build at $0 a month, Ship from $50, Scale from $500, and Enterprise at a custom price. Each plan includes a monthly allowance of every metered resource and charges per unit above it. Allowances reset on the first of the month and don't roll over. On Build, the allowance is a hard cap: requests fail once it runs out, and nothing is billed.
| Meter | Build ($0) | Ship (from $50) | Scale (from $500) |
|---|---|---|---|
| Agent-session minutes | 1,000 | 5,000, then $0.01/min | 50,000, then $0.01/min |
| Concurrent agent sessions | 5 | 20 | From 50, up to 600 |
| LiveKit Inference credits | $2.50 | $5, then model prices | $50, then discounted model prices |
| Inference concurrency | 5 | 20 | 50 |
| WebRTC participant minutes | 5,000 | 150,000, then $0.0005/min | 1.5M, then $0.0004/min |
| Downstream data | 50 GB | 250 GB, then $0.12/GB | 3 TB, then $0.10/GB |
| US local number / inbound minutes | 1 free / 50 min | 1 free, then $1/mo / 100 min, then $0.01/min | 1 free, then $1/mo / 1,000 min, then $0.01/min |
| Third-party SIP minutes | 1,000 | 5,000, then $0.004/min | 50,000, then $0.003/min |
| Session recordings | 1,000 min | 5,000 min, then $0.005/min | 50,000 min, then $0.005/min |
| Observability events | 100,000 | 500,000, then $0.00003 each | 5M, then $0.00003 each |
Source: LiveKit pricing and its plain-text version, which also lists per-model inference prices and the calculator's assumptions.
How the meters count
An agent session is billed per second with a 10-second minimum. Metering starts when the agent connects to the room and stops when the room ends or the agent leaves. An agent that receives a job and never connects is not billed. If your agent lingers after the caller hangs up, the minutes keep counting until you call ctx.shutdown().
LiveKit Inference bills each model type by its own unit. STT is billed per second of connection time, the LLM per token, and TTS per character. So STT runs for the whole session, while the LLM and TTS cost depends on how much the agent says.
Two budgets at 20,000 minutes a month
Both budgets assume 20,000 agent-session minutes on the Ship plan, about 4,000 five-minute conversations. Both use the models from LiveKit's own worked example: Gemma 4 31B, Deepgram Nova-3 and Inworld Realtime TTS 2.0 through LiveKit Inference. With LiveKit's default assumptions of 3,000 input tokens, 175 output tokens and 600 TTS characters per minute, that stack costs $0.0212 per minute ($0.0014 LLM, $0.0048 STT, $0.0150 TTS). Your token and character counts will differ.
| Line | Web app (WebRTC) | Phone app (US local inbound) |
|---|---|---|
| Ship plan base | $50.00 | $50.00 |
| Agent sessions: 15,000 min over the 5,000 included × $0.01 | $150.00 | $150.00 |
| Inference: 20,000 × $0.0212, less $5 credit | $419.00 | $419.00 |
| WebRTC participant minutes: 20,000 of 150,000 included | $0.00 | n/a |
| Downstream data, assumed inside 250 GB | $0.00 | n/a |
| Inbound minutes: 19,900 over the 100 included × $0.01 | n/a | $199.00 |
| One US local number | n/a | $0.00 (first number free) |
| Total without observability | $619.00 ($0.031/min) | $818.00 ($0.041/min) |
| Optional: recordings, 15,000 min × $0.005 | $75.00 | $75.00 |
| Optional: events at 100/min, 1.5M over the 500,000 included × $0.00003 | $45.00 | $45.00 |
| Total with observability | $739.00 | $938.00 |
The data line assumes an audio-only agent stays inside the 250 GB allowance. 100 events per minute is LiveKit's own calculator assumption for observability. Carrier fees, your own servers, and any model you call outside LiveKit Inference are not in these totals.
The two bills differ by the phone line, $199 a month. Everything else is identical, because the agent does the same work either way.
LiveKit's own worked example prices this phone stack at $0.0412 per minute. Multiplied by 20,000 that gives $824, against $818 above. The per-minute figure leaves out the $50 base fee and the included allowances. On Ship above 5,000 minutes those roughly cancel: 5,000 agent minutes ($50), 100 inbound minutes ($1) and $5 of credits offset the base fee with $6 to spare. Below 5,000 minutes the base fee buys allowance you don't use, and the per-minute figure understates the bill.
If you bring your own SIP trunk instead of a LiveKit number, LiveKit charges $0.004 per minute above 5,000 included, so 15,000 × $0.004 = $60. Your carrier bills separately. LiveKit's calculator notes suggest about $0.008 per minute for Twilio or Telnyx, roughly $160 here. That makes $220 against $199 for LiveKit's own number.
When concurrency forces Scale
Ship allows 20 concurrent agent sessions and 20 concurrent inference connections. A support line with a busy hour above 20 callers needs Scale, whatever its monthly minutes. Scale changes the arithmetic. Its 50,000 included agent minutes cover this whole workload, its inference prices are lower (Nova-3 $0.0042, Inworld TTS 2.0 $15 per million characters), and it includes $50 of credits.
| Line | Web app | Phone app |
|---|---|---|
| Scale plan base | $500.00 | $500.00 |
| Agent sessions (inside 50,000 included) | $0.00 | $0.00 |
| Inference: 20,000 × $0.0146, less $50 credit | $242.00 | $242.00 |
| Inbound minutes: 19,000 over 1,000 × $0.01 | n/a | $190.00 |
| Total without observability | $742.00 | $932.00 |
At this volume Scale costs about $120 a month more than Ship and buys the concurrency headroom. At 50,000 minutes the order flips. Ship would charge $450 in agent-session overage alone, and Scale includes those minutes.
The calculator applies Ship rates and allowances: $50 base, 5,000 agent minutes, $5 credits, 100 US local inbound minutes. It leaves out observability, data transfer above 250 GB and carrier fees.
Does $0.01/min include the LLM?
No. The $0.01 agent-session rate pays for LiveKit hosting your agent process while it sits in a room. STT, LLM and TTS are billed either as LiveKit Inference usage or by the provider directly when you use a plugin with your own key. In the budgets above, models cost more than twice the agent-session line. A speech-to-speech model replaces the STT and TTS lines: LiveKit's calculator lists OpenAI GPT Realtime at an assumed $0.0676 per minute.
What costs extra?
Beyond agent sessions and inference, the pricing page lists separate charges for phone numbers and inbound minutes (toll-free numbers are $2 a month plus $0.02 a minute), SIP and Twilio or WhatsApp connector minutes, WebRTC minutes and data above the allowance, session recordings, observability events, and voice isolation above its allowance ($0.0012 a minute on Ship and Scale). Private links on Scale cost $50 each plus $0.10 per GB. Text and audio simulations list $0.02 and $0.05 per turn, free through October 31 on paid plans.
Agent-session concurrency vs. inference concurrency
LiveKit counts them separately. Agent-session concurrency is the number of live conversations your deployed agents hold at once. Inference concurrency is the number of simultaneous connections to LiveKit Inference per model type. Ten calls using LiveKit's STT open ten STT connections. On Ship both limits are 20. If you call your models directly with your own keys, inference concurrency no longer applies, and your providers' rate limits take its place.
Self-hosting LiveKit
The LiveKit server and the Agents framework are open source, and you can run both yourself. That removes LiveKit's agent-session meter and replaces it with your servers and your on-call time. A self-hosted agent that joins rooms on LiveKit Cloud still counts against WebRTC participant minutes. Agent insights do not work with self-hosted media servers. Pipecat can run in LiveKit rooms too; our Pipecat vs. LiveKit note covers that setup.
The model line is usually the one worth comparing across vendors. Duplex Compute is a full-duplex voice API with tool calling at $0.01 per connected minute ($0.60 an hour). We have not published a LiveKit plugin, and a session rate does not make the whole agent cheaper on its own. Count the full bill.
Related notes
- Voice agent pricing: calculator and billing formulas
- Pipecat vs. LiveKit
- Vapi alternatives for developers
- Retell AI pricing
- Self-hosted voice AI vs. managed APIs
Pricing a voice agent at volume?
If the model lines dominate your LiveKit bill, tell us your expected voice hours, the tools your agent calls and your current provider. We'll tell you plainly whether a $0.01-per-minute full-duplex session fits.
Request access