Every "LiveKit vs. Retell vs. Vapi" post you'll find, including the one LiveKit itself publishes, tells you LiveKit is the cheaper option. At 1,000 monthly minutes, LiveKit runs $10 to $73 against $70 to $310 for Retell and $70 to $116 for Vapi. At 50,000 minutes, the gap widens to $1,096-$3,844 versus $3,500-$15,500 and $3,492-$5,782. The numbers are real and the source is primary. The framing around them is still misleading, because every one of those figures prices LiveKit Cloud, the managed product, not the open source server running on your own infrastructure that most people mean when they say "self-host LiveKit."
Three products share one name
Comparisons of LiveKit voice AI agents against Retell, Vapi, and Bland keep running into the same mix-up.
"LiveKit" means three different financial commitments depending on which one you're actually comparing:
- LiveKit Cloud, the managed service, billed on the Build, Ship, and Scale plans: Build is free with 1,000 agent session minutes included and $0.01 a minute after that; Ship starts at $50 a month with 5,000 minutes included; Scale starts at $500 a month with 50,000 minutes included and up to 600 concurrent sessions. Enterprise pricing is custom.
- LiveKit the open source project, self-hosted on your own servers, with no plan fee and no vendor concurrency cap at all.
- LiveKit Agents, the open source framework for building the voice pipeline itself, which runs on top of either option above and still calls out to whichever speech-to-text, language model, and text-to-speech providers you choose.
The comparison tables circulating right now, LiveKit's own included, measure the first item. They almost never put a number on the second.
What LiveKit Cloud actually buys you
The Build-Ship-Scale structure explains why LiveKit Cloud undercuts Retell and Vapi in a straight cost comparison: Retell and Vapi bill a single per-minute rate on every call with no free tier, while LiveKit Cloud bundles a block of included minutes into a monthly plan fee and only charges the $0.01 overage rate past that, according to LiveKit's pricing page. At low volume, the free Build plan can cost nothing. At high volume, the per-minute overage rate stays far below what Retell or Vapi charge per minute outright.
That is still a managed product. LiveKit runs the servers, the TURN relays, and the scaling. You are renting infrastructure at a lower markup than Retell or Vapi, not avoiding infrastructure cost altogether. The plan tiers also cap how many calls can run at once, not just how many minutes you use: Build allows 5 concurrent sessions, Ship allows 20, and Scale allows up to 600, per LiveKit's pricing page. A business running one receptionist line rarely bumps into that ceiling. A multi-location deployment handling simultaneous calls across several sites can hit the concurrency cap before it hits the minutes cap, which changes which plan actually fits.
What self-hosting actually requires
LiveKit's own self-hosting documentation lists what a production deployment needs: a domain name with a trusted SSL certificate, a load balancer or reverse proxy terminating HTTPS, specific ports open for the media server (7880 for HTTP, 7881 for RTC, a UDP range from 50000 to 60000), a TURN relay on port 5349 for callers behind restrictive networks, and a Redis instance recommended for production. The same documentation states plainly that "scalability of LiveKit is bound by CPU and bandwidth," not GPU, since the server is routing audio streams rather than running a model.
That last part is the piece every comparison skips, and it's the same blind spot we flagged when breaking down what AI automation projects actually cost: a missing line item doesn't mean the cost is zero, it means nobody priced it yet.
The provider bill doesn't change based on who hosts the server
Here's the detail that gets lost once the Cloud-vs-self-hosted framing takes over: speech-to-text, the language model, and text-to-speech are billed by Deepgram, OpenAI, ElevenLabs, or whichever providers you pick, the same way whether you're running on LiveKit Cloud, self-hosted LiveKit, or a fully managed platform like Retell or Vapi. LiveKit's own comparison folds those provider costs into its all-in totals for every platform. As a rough illustration, subtracting LiveKit Cloud's cited $1,096-$3,844 range from Retell's cited $3,500-$15,500 range at that same 50,000-minute volume puts the gap at roughly $2,400 to $11,600 a month; the same subtraction against Vapi's narrower $3,492-$5,782 range puts its gap at roughly $1,900 to $2,400 a month. Neither gap is a difference in AI model cost, since the same provider calls are already folded into both sides of the subtraction. It's the orchestration markup each platform charges on top of a provider bill that stays roughly constant no matter who's hosting the pipeline.
Self-hosting removes that orchestration markup down to zero. It does not touch the provider bill at all, and it adds your own infrastructure cost and ops time in its place.
When each option actually fits
- Fully managed (Retell, Vapi, Bland): the right choice below a few thousand monthly minutes, or when a team wants zero infrastructure to think about and will pay the per-minute premium for that. Our retell vs. vapi vs. bland comparison covers how those three stack up against each other specifically.
- LiveKit Cloud: the middle option once volume climbs into the tens of thousands of minutes a month and the per-minute savings start to outweigh a $50-$500 plan fee, without taking on any server management.
- Self-hosted LiveKit: worth evaluating only once monthly volume is high enough, or a data-residency requirement is strict enough, that the eliminated plan fee and per-minute overage clearly outweigh running your own SFU, Redis, and TURN relay, plus whoever's time it takes to keep all three patched and online.
That last tradeoff is infrastructure work, not a voice AI decision, which is exactly the kind of build our workflow automation team scopes before anyone touches a voice pipeline: what it costs to run the thing yourself, measured against what it costs to keep renting it from someone else.
The honest version of "LiveKit is cheaper"
LiveKit Cloud being cheaper than Retell and Vapi, at the volumes LiveKit itself publishes, is true and worth taking seriously if you're already shipping tens of thousands of minutes a month. "Self-hosting LiveKit is cheaper still" is the part that's unproven, not because it's false, but because the party with every incentive to publish that number, LiveKit, chose not to. A server bill, a bandwidth bill, and a few hours a month of someone watching Redis and renewing a TLS certificate are real costs with no published range to compare against.
That gap between "cheaper on paper" and "cheaper once someone's time is on the server" is the whole story here, and it's the same gap our voice AI team walks every client through before a single line of a build quote gets written: the comparison post gave you two real numbers and one honest blank, and the blank is the one your invoice will actually reflect.
