A LiveKit budget depends on where your agent runs, how users connect, and which services process each conversation. For engineering teams planning production voice agents, the useful number is a workload subtotal with clear boundaries. Separate the Cloud bill from model, carrier, infrastructure, and operating costs before comparing architectures.
How much does LiveKit cost?
LiveKit Cloud has a free Build plan, Ship starting at $50 per month, Scale starting at $500 per month, and custom Enterprise pricing. Paid plans include resource-specific allowances, then charge for excess usage. The $0.01 hosted-agent minute is one line of a voice-agent bill. Models, the user’s connection channel, recordings, events, and data can add separate charges.
All dollar rates and plan allowances below are in USD as of November 4, 2026. The worked examples are illustrative Cloud subtotals, not all-in production prices.
Start with three decisions:
- Hosting: Does LiveKit host the agent, just the media connection, or neither?
- Channel: Does the user connect through a browser, a LiveKit phone number, or a third-party carrier?
- Inference: Are speech and language models billed through LiveKit or through your own provider accounts?
Those choices determine which meters belong in the calculation. Adding every published rate to every conversation minute will overstate some workloads and miss costs in others.
Choose the deployment boundary first
LiveKit provides both open-source software and managed Cloud services. Self-hosting an agent does not necessarily mean self-hosting the media network.
WebRTC is the real-time connection used by browser clients and agent participants. SIP, or Session Initiation Protocol, connects telephone calls through a carrier trunk.
| Deployment | LiveKit Cloud resources to budget | Costs and work outside that bill |
|---|---|---|
| Cloud-hosted agents and Cloud media | Plan, hosted-agent time, connection channel, selected inference, data, optional recording and events | Application backend, integrations, external storage and services, carrier if using third-party telephony |
| Your agent servers with Cloud media | Plan, WebRTC participant time including your agents, channel, data, selected inference and observability | Agent compute, warm capacity, deployment, autoscaling, monitoring and incident response |
| Your agent servers and media servers | Only Cloud services you actually consume | Agent/media compute, networking, telephony services, recording services, model providers, storage and operations |
Cloud agent hosting includes builds, autoscaling, and load balancing. A Cloud deployment should not be budgeted as though your team must operate its own media servers.
In the hybrid option, remove the hosted-agent session charge. Your agent instead counts as a connected WebRTC participant on Cloud media. One browser user plus one self-hosted agent creates two participant clocks. A self-hosted agent connected to a phone session can also consume WebRTC participant time alongside the phone channel. For a Cloud-hosted agent, use the hosted-agent meter without adding a second WebRTC charge for that same agent.
Fully self-hosted media removes Cloud transport charges for the resources you no longer use. It replaces them with your infrastructure and operating responsibilities. The software’s open-source availability does not make a production deployment cost-free.
LiveKit Cloud plans and included usage
Each row is a separate allowance in LiveKit’s Cloud pricing. Five thousand included agent minutes cannot pay for five thousand additional recording minutes after the recording allowance is exhausted.
| Resource per month | Build | Ship | Scale |
|---|---|---|---|
| Plan entry price | $0 | $50 | $500 |
| Hosted-agent minutes | 1,000 | 5,000, then $0.01/min | 50,000, then $0.01/min |
| Inference credit | $2.50 | $5 | $50 |
| WebRTC participant minutes | 5,000 | 150,000, then $0.0005/min | 1,500,000, then $0.0004/min |
| Downstream data | 50 GB | 250 GB, then $0.12/GB | 3 TB, then $0.10/GB |
| Agent audio 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/event | 5,000,000, then $0.00003/event |
| Third-party SIP minutes | 1,000 | 5,000, then $0.004/min | 50,000, then $0.003/min |
Build is capped. It does not automatically convert exhausted allowances into paid overages. Relevant new requests fail once a limit is reached. Free allowances are shared across a user’s free projects, with separate per-project development allowances for turn detection and interruption handling.
Included allowances reset on the first day of each calendar month and do not roll over. Paid plans meter excess usage by resource. Shared plans across projects are an Enterprise feature, so separate paid projects should not be treated as one pooled allowance.
Inference credit is a dollar balance shared across supported models. It is not a fixed entitlement to a particular number of conversation minutes.
For each paid-plan resource, calculate max(0, monthly billable usage - included allowance) × rate. Add those results to the plan price. Deduct inference credit from model dollars, rather than from duration.
Capacity can determine the plan before volume does
Build supports five concurrent hosted-agent sessions; Ship supports 20. Scale starts at 50, with increases up to 600 requested through the dashboard. The higher limit is not automatically provisioned.
Agent concurrency, model-service limits, and total network connections are different constraints. A project can have monthly minutes remaining while new work fails at a relevant concurrency or request limit. Large language models also have request and token limits; one headline concurrency number cannot describe the whole pipeline.
Enterprise pricing, volume discounts, shared allowances, support terms, and contracted capacity are quote-specific. Enterprise has an annual commitment. A rate-card uptime figure is not a substitute for the service terms of a particular contract.
Understand what each meter counts
Conversation duration is a useful planning input. It is not the billing unit for every service. Agent, participant, and speech-service connections can begin and end at different points in the same session.
| Meter | What counts | Common budgeting mistake |
|---|---|---|
| Hosted agent | Time connected to the room, until the room ends or the agent disconnects | Counting only speaking time |
| WebRTC participant | Each applicable participant’s connected time | Counting one room instead of its participants |
| Third-party SIP | Each caller’s connected trunk time | Treating it as the carrier’s phone-network bill |
| Speech-to-text (STT) | Connection time through LiveKit Inference | Excluding silence while STT remains connected |
| Large language model (LLM) | Input and output tokens, with model-specific cache units | Counting only newly spoken user words |
| Text-to-speech (TTS) | Generated text characters through LiveKit Inference | Assuming characters equal connected minutes |
| Observability | Recorded audio duration and individual events | Assuming recordings include unlimited transcripts, traces, and logs |
Hosted-agent, WebRTC, and third-party SIP sessions have a 10-second minimum and then one-second increments. A two-second hosted-agent connection therefore bills ten seconds. Recorded agent audio also has a 10-second minimum and one-second increments. Native LiveKit-number inbound time is metered from the first second.
Rounding happens before monthly aggregation. For example, downstream usage of 0.001 GB rounds to 0.01 GB, while 0.011 GB rounds to 0.02 GB. Rounding only the monthly byte total can undercount the bill. The resource metering rules define the individual clocks and increments.
An agent that never connects to the room does not incur hosted-agent time. An agent that remains connected after the useful conversation ends can keep consuming it. Session cleanup belongs in both the implementation and the cost model.
Add model usage in its native units
A cascaded voice pipeline uses STT, an LLM, and TTS. For LiveKit-routed inference, calculate each component’s usage, add the dollar amounts, then deduct the plan’s inference credit once.
These selected LiveKit Inference rates support the examples. They are not direct-provider tariffs or recommendations about model quality. In the tables, M means million.
| Model | Ship | Scale |
|---|---|---|
| Gemma 4 31B, uncached input | $0.40/1M tokens | $0.40/1M tokens |
| Gemma 4 31B, output | $1.20/1M tokens | $1.20/1M tokens |
| Deepgram Nova-3 Monolingual STT | $0.0048/connection min | $0.0042/connection min |
| Inworld Realtime TTS 2.0 | $25/1M characters | $15/1M characters |
LLM input includes repeated conversation context, system instructions, and tool content sent on each request. Retries, fallbacks, caching, and unused speculative generations can change the total. A universal tokens-per-call assumption hides those differences.
Provider plugins can use your own accounts and API keys. In that case, the model bill sits outside LiveKit. Mixed routing is possible, so allocate each model invocation to its actual billing route once.
A realtime speech-to-speech model replaces the cascade. Its bill may use audio/text tokens, duration, or other model-specific units. Do not automatically add a second STT and TTS charge unless your implementation actually invokes those services separately.
Phone costs depend on the connection path
Native LiveKit phone numbers currently support US inbound calls only. Third-party SIP supports inbound and outbound calling through your carrier.
| Phone resource | Ship | Scale |
|---|---|---|
| US-local LiveKit number | One included, then $1/number/month | One included, then $1/number/month |
| Native US-local inbound | 100 min included, then $0.01/min | 1,000 min included, then $0.01/min |
| US toll-free LiveKit number | $2/number/month | $2/number/month |
| Native US toll-free inbound | $0.02/min | $0.02/min |
| Third-party SIP | 5,000 min included, then $0.004/min | 50,000 min included, then $0.003/min |
For native inbound, use the native inbound meter and applicable number rental. Do not add a third-party SIP fee to the same call simply because telephony uses SIP internally.
For a third-party trunk, add the actual carrier bill: number rental, inbound or outbound calling, destination, additional transfer legs, rounding, taxes, and fees. LiveKit’s SIP rate alone does not price a completed phone call.
Native phone-number forwarding through TransferSIPParticipant is not currently supported. That restriction belongs to LiveKit Phone Numbers; it should not be generalized to all third-party SIP configurations.
Worked example: 40,000 voice minutes on Ship
Illustrative scenario assumption. Assume one project handles 10,000 four-minute sessions in a calendar month, using Cloud-hosted agents and one browser user per session.
Also assume:
- Agent, browser, and STT clocks each total exactly 40,000 minutes, with no short sessions, observers, extra agents, or transfers.
- Each conversation minute uses 1,000 uncached Gemma 4 31B input tokens, 100 output tokens, and 300 Inworld Realtime TTS 2.0 characters. STT uses Nova-3 Monolingual throughout.
- Billed downstream data totals 100 GB after individual rounding.
- All audio is recorded, with 50 billable observability events per minute, or 2 million events.
- No other usage or paid add-ons consume the project’s allowances.
The token, character, and event counts are planning inputs, not measured industry averages.
Calculate inference first
| Component | Calculation | Cost |
|---|---|---|
| LLM input | 40M tokens × $0.40/1M | $16.00 |
| LLM output | 4M tokens × $1.20/1M | $4.80 |
| STT | 40,000 min × $0.0048 | $192.00 |
| TTS | 12M characters × $25/1M | $300.00 |
| Inference after credit | $512.80 − $5 | $507.80 |
Apply each remaining allowance separately
| Cloud line | Calculation | Cost |
|---|---|---|
| Ship plan | Monthly entry price | $50.00 |
| Hosted agent | (40,000 − 5,000) × $0.01 | $350.00 |
| Browser participant | 40,000 within 150,000 min | $0.00 |
| Downstream data | 100 within 250 GB | $0.00 |
| Inference | From the calculation above | $507.80 |
| Audio recording | (40,000 − 5,000) × $0.005 | $175.00 |
| Events | (2M − 500,000) × $0.00003 | $45.00 |
| Illustrative Cloud subtotal | Sum of billed lines | $1,127.80 |
The browser and data lines have no usage overage under these assumptions. They still consume their respective allowances.
Change only the user’s channel
Keep the same hosted-agent, model, recording, and event usage. Replace the browser connection with one phone caller per session:
| Connection path | Additional Cloud channel charge | Cloud subtotal |
|---|---|---|
| Browser baseline | $0 overage | $1,127.80 |
| Native US-local inbound, one included number | (40,000 − 100) × $0.01 = $399 | $1,526.80 |
| Third-party SIP | (40,000 − 5,000) × $0.004 = $140 | $1,267.80 + carrier |
The two phone rows are alternatives. Their charges are not added together. The native-number row cannot price outbound or non-US calling, and the SIP subtotal excludes the carrier.
These calculations also exclude taxes and fees, application/backend hosting, databases, tools, external monitoring, archival storage, development, and operations. They include no media import/export, simulations, voice-isolation overage, or other paid add-ons. The subtotal is not a universal per-minute offer.
A larger Scale workload uses different allowances
Illustrative scenario assumption. Assume 25,000 four-minute Cloud-hosted agent sessions on Scale, each with one third-party SIP caller. Agent, caller, and STT clocks each total 100,000 minutes. Use the same per-minute model inputs, record all audio, and generate 50 events per minute. Assume 200 billed GB of downstream data, no other usage or add-ons, and peak demand within the project’s confirmed limits.
| Cloud line | Calculation | Cost |
|---|---|---|
| Scale plan | Monthly entry price | $500.00 |
| Hosted agent | (100,000 − 50,000) × $0.01 | $500.00 |
| Third-party SIP | (100,000 − 50,000) × $0.003 | $150.00 |
| Inference after credit | $52 LLM + $420 STT + $450 TTS − $50 | $872.00 |
| Recording | (100,000 − 50,000) × $0.005 | $250.00 |
| Events | 5M within 5M allowance | $0.00 |
| Downstream data | 200 GB within 3 TB | $0.00 |
| Illustrative Cloud subtotal | Before carrier and excluded production costs | $2,272.00 |
Monthly volume alone does not establish the necessary concurrency. Nor does this example prove Scale is the cheapest plan for another workload. Allowances, model discounts, required features, and peak capacity all affect that decision.
Recording, exports, and testing add separate meters
Agent observability records audio inside the agent and uploads it to LiveKit. Media Egress, LiveKit’s recording/export service, exports media to your own storage. They are separate pipelines, with separate costs.
Agent insights supports Cloud-hosted agents and self-hosted agents connected to Cloud media. It does not support fully self-hosted media. Default retention is 30 days, so an included recording allowance should not be treated as an indefinite archive.
Other cost lines include:
- Transcoding for import/export: Ship includes 600 minutes and Scale 8,000, shared between stream import and recording/export. Audio-only excess costs $0.005/min on Ship or $0.004/min on Scale; video costs $0.02 or $0.015/min respectively.
- Track egress: A separate allowance of 600 minutes on Ship or 8,000 on Scale, then $0.001/min.
- Voice isolation: Ship includes 1,000 minutes and Scale 10,000, then $0.0012/min.
- Simulations: On Ship and Scale, text turns cost $0.02 and audio turns $0.05. Model/provider spend can also depend on the simulation configuration.
Own-bucket storage, requests, transfer, and long-term retention remain separate. A free transcode-less import does not make its downstream network usage free.
What self-hosting actually costs
For a hybrid deployment, retain the Cloud resources you consume and add your agent infrastructure. For fully self-hosted media, add the media and supporting services as well.
Budget these separately:
- Compute and capacity: Agent server-hours, media servers, warm/idle instances, peak load, failover, and separate environments.
- Network services: Bandwidth, load balancers, domains, certificates, TURN network relays for clients whose direct connection paths fail, and Redis where required by the architecture.
- Optional services: SIP, Ingress for media import, Egress for export, object storage, and backups.
- Operations: Autoscaling, long-running job draining, upgrades, telemetry, alerts, incident response, and evaluation work.
- External dependencies: Models, carriers, application tools, and databases not already included elsewhere.
Provisioned server-hours and measured utilization drive infrastructure cost. Conversation minutes alone cannot determine an instance type or staffing budget. The self-hosted deployment requirements explain the agent lifecycle and operating responsibilities.
Self-hosted agents can also incur LiveKit model-request charges. On Ship and Scale, audio turn detection costs $0.50 per 1,000 requests beyond 7,500 development requests per project each month. Adaptive interruption handling costs $1.20 per 1,000 requests beyond 40,000. A user turn can trigger multiple requests. Cloud-hosted agents include these services without an extra request charge.
There is no defensible universal break-even point between Cloud and self-hosting without the workload, architecture, effective cloud rates, and operating effort.
When a managed conversational runtime fits
LiveKit is a legitimate choice when you want to build with its framework and control the agent application. LiveKit Cloud also removes substantial hosting work. The remaining decision is how much conversational application and operating responsibility your team wants to retain.
Our managed production platform helps technical teams build and run production voice AI agents through a managed runtime, APIs, and a web application. We provide agent configuration with prompts, models, voices, and tools; phone and embedded web channels; and webhooks. Your team still owns its application behavior, integrations, and channel configuration.
For completed calls, our Call Inspector brings together transcripts, recordings, model interactions, tool events, and timeline analysis. That is a concrete operating workflow to compare alongside your hosting and provider choices.
Keep the same workload and cost boundaries when evaluating a different architecture. Our voice-agent pricing guide explains the broader platform comparison. To explore our managed approach, start with Dasha’s documentation.
