The circulating “switch at ~10,000 minutes a month” rule is an infrastructure-only number. Load the engineering and the crossover moves to 90,000–183,000 minutes a month; strip it out, as most comparison tables do, and it falls to between 360 and 8,100 depending on which LiveKit Cloud tier you buy. Every price was read on the vendor’s own pricing page on 18 September 2026.
The managed platforms — Vapi, Retell AI, Bland AI — publish a per-minute price. The frameworks — LiveKit Agents and Pipecat — are free to download and expensive to own, and the part that costs money is not on any pricing page. The decision turns on which of those two numbers you are comparing.
Three adjacent questions are answered elsewhere on this site and are not repeated here: which framework, in our architectural comparison of LiveKit Agents and Pipecat; how much build work there is, priced in engineer-days, in how long it takes to build a voice AI agent in-house; and what a minute of model time costs by architecture, in our breakdown of voice AI cost per minute, realtime versus pipeline. This page is the crossover between them.
What the 10,000-minute rule actually claims
The threshold is stated plainly and testably. Hamming AI’s voice agent stack guide, read 18 September 2026, lists as a key takeaway: “Platform selection: <10K mins/month or <1 month timeline = managed platform; >10K mins + engineering team = custom build”, and in its FAQ: “Build (LiveKit, Pipecat) when: you exceed 10-50K minutes/month (80% cost savings possible)”.
The same guide, published 23 December 2025, carries a higher number elsewhere on the page. Its stack-selection table gives “Cost efficiency at scale (>50K mins/month)” as the custom-build trigger, and its FAQ states both “For >50K mins/month, consider custom builds with LiveKit for 80% cost reduction” and “At >50K mins/month, custom builds achieve 80% savings vs managed platforms”. It also says of its own recommendations that “These are starting points, not gospel”. The 10,000 figure is the one that circulates; the 50,000 figure is the one that travels with the caveats, and it is much closer to what follows.
Credit where it is due before anything is disputed. Its cost figures are close: it puts Vapi at $0.05–$0.13 per minute all-in and $500–$1,300 a month at 10,000 minutes. The top of that range is almost exactly what Vapi’s own calculator returns for its default components ($1,318); the bottom is the $0.05 hosting fee standing alone, before any model cost, which the same calculator never produces. And it does not hide the engineering — its platform table lists the custom option as “~$800 + eng. time” and its FAQ says “2+ engineers for 3+ months initially”.
The engineering is named but never priced, so it never enters the column being compared. That is not an error; it is an invitation to do the arithmetic yourself. So we did.
The criteria that decide this before any price does
Price is the fourth question, not the first. Stop at the first row that describes you, because a hard constraint outranks a cost curve.
| Criterion | If this is true of you | The price comparison is irrelevant because |
|---|---|---|
| Ownership of a long-running audio service | Nobody on the team will carry a pager for a process that holds live phone calls | A framework is a codebase you operate. There is no version where you own LiveKit Agents or Pipecat and nobody is on call. |
| Call-rate ceilings | You need more than 2,000 calls a day or more than 1,000 in an hour | Bland AI publishes those exact caps on its Build plan. Ceilings are a feature question, not a cost question. |
| SIP registration | Your carrier or PBX requires the agent to REGISTER rather than accept a static trunk | The LiveKit telephony feature matrix lists SIP Registration (REGISTER) as Not Supported, as at 18 September 2026. |
| Regulated data handling | You need a signed BAA for HIPAA-eligible handling | Vapi publishes a $2,000 per month add-on; Bland lists it in its Enterprise tier. Fixed costs that do not scale with minutes. |
| None of the above | You have the team, the time and no hard ceiling | Now it is a cost question, and the calculation below is the one to run. |
The worked crossover calculation, end to end
The method: compare each managed platform at its published rate against the same volume run on LiveKit Cloud, with the engineering load added only to the framework side, because only the framework side incurs it.
Published inputs (vendor pages, read 18 September 2026)
| Vendor | What the page states | Per minute |
|---|---|---|
| Vapi | “Usage is your Vapi hosting fee ($0.05/min) plus the model costs for each minute your agents run.” Models are billed “at the provider’s listed price with no markup”. The Core success package is $29/month; a usage-only tier with no package is also offered | $0.05 platform, plus models, plus $29/month Core |
| Vapi (models) | Calculator component ranges: Deepgram $0.0095–$0.0099, OpenAI $0.0077–$0.0452, ElevenLabs $0.0146–$0.0238 | $0.0318–$0.0789 |
| Retell AI | “$0.07-$0.31 / min for AI Voice Agents”. Calculator default breaks down as Retell Voice Infra $0.055, LLM $0.04, TTS $0.015 | $0.11 (default config) |
| Bland AI | Start $0.14/min with $0 platform fee; Build $0.12/min with a $299/month platform fee. LLM, STT and TTS are stated as included in the per-minute rate | $0.12 plus $299/month |
| LiveKit Cloud (Ship) | $50/month includes 5,000 agent session minutes then $0.01/min, and 5,000 third-party SIP minutes then $0.004/min | $0.014 above allotment |
| LiveKit Cloud (Scale) | $500/month includes 50,000 agent session minutes then $0.01/min, and 50,000 third-party SIP minutes then $0.003/min | $0.013 above allotment |
| LiveKit Inference | GPT-5.6 Luna $0.0008, Deepgram Nova-3 Monolingual $0.0048, Deepgram Aura-2 $0.0180 (Build/Ship list) | $0.0236 cascade |
| Pipecat Cloud | agent-1x (0.5 vCPU, 1 GB) $0.01/min active, $0.0005/min reserved; unlimited concurrency; LLM, STT and TTS “Billed to provider” | $0.01 platform |
That table is doing more work than the headline rates. Vapi and Pipecat Cloud sell orchestration and pass the models through; Bland sells a bundle. Retell publishes both, which is why its Voice Infra line of $0.055 a minute is directly comparable to Vapi’s $0.05 and Pipecat Cloud’s $0.01, and its $0.11 headline is not.
Which LiveKit tier you buy moves the framework line more than any model choice does. The framework column below takes whichever tier is cheaper at that volume: Ship up to about 37,100 minutes a month, Scale above it, where the $500 fee is repaid by the larger allotments. At 5,000 and 10,000 minutes that is Ship at $50; at 50,000 and 200,000 it is Scale at $500. Substitute the other tier and every number in the bottom row of the threshold table changes by an order of magnitude.
The assumptions, stated as assumptions
Everything below this line is ours, not a vendor’s, and none of it is a measurement. Substitute your own and the structure survives.
- A1. Loaded engineer cost: US$1,000 per engineer-day. Roughly a US$180,000 total employment cost over about 180 productive days a year. This is the single input most worth replacing with your own.
- A2. Build: 83 engineer-days, amortised straight-line over 24 months = $3,458 a month. That figure is our own published estimate for an outbound-at-volume profile, with the stage-by-stage table behind it on our in-house build timeline page. It is an estimate, not a measurement.
- A3. Evaluation and model-churn maintenance: 3 engineer-days a month = $3,000. Regression suites, plugin migrations, prompt re-testing after a model changes underneath you.
- A4. On-call: 2 engineer-days a month = $2,000, covering the roster allowance and the interrupted work it causes.
- A5. All minutes are outbound over a third-party SIP trunk, four-minute average call, trunk cost excluded on both sides because it is the same either way.
- A6. 22 business days a month and a 12-hour dialling window, so 15,840 wall-clock minutes per line. Everything above that is concurrency: 200,000 minutes a month is about 13 lines talking continuously through the window, and more than that at peak. Concurrency charges are excluded on both sides, and that is not quite symmetrical. LiveKit Cloud includes 20 concurrent agent sessions on Ship and 50 on Scale at no extra charge, Bland includes 50 on Build, Retell includes 20 on pay-as-you-go and states no rate above that on its pricing page, and Vapi includes 10 on Core and charges $10 per line per month beyond it. Vapi is the one published exception, worth a few hundred dollars a month at 200,000 minutes, and counting it would move the crossover against Vapi down, not up.
- A7. Currency. None of the five pricing pages prints a currency code beside a price in its visible tables. Vapi and Retell declare
"priceCurrency": "USD"in schema.org markup, and Twilio carriesdata-currency-code="usd". On the LiveKit, Daily and Bland pages we could locate no currency code in either the visible text or the page source, read 18 September 2026. Read every “$” as unconverted. - A8. Inference is priced from the LiveKit Build/Ship model list throughout, including in the rows that buy Scale, where the same three models are cheaper ($0.0212 rather than $0.0236 a minute). That overstates the framework cost slightly and therefore overstates the crossovers in the top row of the threshold table by about 5 per cent; on the Scale inference list they are 173,900, 109,600 and 93,300 rather than 183,200, 113,200 and 96,000. The conclusion does not turn on it.
A2 plus A3 plus A4 is $8,458 a month. That is the number the whole argument turns on, and it is entirely assumption.
The result
| Minutes/month | Vapi all-in | Retell | Bland Build | Framework infra only | Framework + engineering |
|---|---|---|---|---|---|
| 5,000 | $438–$674 | $550 | $899 | $168 | $8,626 |
| 10,000 | $847–$1,318 | $1,100 | $1,499 | $356 | $8,814 |
| 50,000 | $4,119–$6,474 | $5,500 | $6,299 | $1,680 | $10,138 |
| 200,000 | $16,389–$25,809 | $22,000 | $24,299 | $7,170 | $15,628 |
The gap between the last two columns is the entire finding. On infrastructure alone the framework is 56% to 81% cheaper at every volume tested — the circulating “up to 80%” saving is real and we reproduce it. Add the engineering and the managed platform stays ahead until somewhere between 90,000 and 183,000 minutes a month.
Does the 10,000-minute rule survive? Only without engineering
The threshold table, solved three ways. The only thing that changes between rows is how much of the engineering you count.
| What you count on the framework side | vs Vapi (cheapest models) | vs Retell $0.11 | vs Bland Build $0.12 |
|---|---|---|---|
| Build, evaluation and on-call (A2+A3+A4) | 183,200 min/mo | 113,200 min/mo | 96,000 min/mo |
| Amortised build only, team already paid for | 72,600 min/mo | 45,800 min/mo | 38,000 min/mo |
| Nothing — infrastructure only, on the Ship tier | 360 min/mo | 580 min/mo | any volume, because the $299 platform fee alone exceeds Ship at $50 |
| Nothing — infrastructure only, buying Scale from day one | 8,100 min/mo | 5,800 min/mo | 2,100 min/mo |
The bottom two rows are where the rule of thumb comes from, and they reconcile it rather than refute it. Compared purely on published infrastructure prices — which is the comparison the circulating tables actually perform — the answer swings by more than an order of magnitude on the LiveKit plan alone. A team buying Scale from day one and comparing against Vapi at its cheapest model set crosses over at 8,100 minutes a month, and Hamming’s own “~$800 + eng. time” infrastructure figure puts that same crossover at about 9,400. So 10,000 is a defensible answer to one question: at what volume does framework infrastructure become cheaper than platform infrastructure, counting no engineering at all. It is not an answer to the question most buyers think they are asking, which is when the whole thing becomes cheaper to own. Count the amortised build and the number moves to 38,000 or more; count evaluation and on-call as well and it moves past 96,000. There is one more assumption set worth naming, because it is the one that makes 10,000 defensible even with engineering in the column.
The Half-Engineer-Day Test
Invert it. What monthly engineering budget would make 10,000 minutes the true crossover?
| Compared against | Implied engineering budget at 10,000 min/mo | At US$1,000 per engineer-day |
|---|---|---|
| Vapi, cheapest model set | $491/month | 0.49 engineer-days a month |
| Retell at $0.11/min | $744/month | 0.74 engineer-days a month |
| Vapi, dearest model set | $962/month | 0.96 engineer-days a month |
| Bland Build at $0.12/min | $1,143/month | 1.14 engineer-days a month |
The Half-Engineer-Day Test: 10,000 minutes a month is the correct switching threshold for a team whose entire framework deployment — build amortisation, evaluation maintenance and on-call included — costs about half an engineer-day a month to own. Price your own ops honestly against that figure. If it is higher, your crossover is higher, and the rule is pointing you at a migration you cannot yet afford.
Half an engineer-day is four hours a month covering everything: the amortised build, every model migration, every regression re-run, and the pager. On the Scale plan, where framework infrastructure at 10,000 minutes is $736 rather than $356, the implied budgets are tighter still at $111 to $763 a month. Apply the test to your own team before you accept the threshold.
What LiveKit native SIP changes about the arithmetic
Older comparisons priced the framework path with a media bridge in the middle: a Programmable Voice call plus Twilio Media Streams forwarding audio over a websocket. The Twilio US voice pricing page, read 18 September 2026 and marked “Pricing current as of August 2026”, lists Media Streams at $0.0044 a minute and ConversationRelay at $0.07 a minute.
LiveKit now carries SIP itself. Its telephony documentation states that LiveKit telephony handles inbound and outbound calls and supports integration with third-party SIP providers, and LiveKit Cloud prices third-party SIP minutes at $0.004 a minute on Ship and $0.003 on Scale, 50,000 included in the Scale allotment. The bridge leg leaves the stack; a trunk remains.
Read the feature matrix rather than the headline, because it is specific about the edges. As at 18 September 2026 it lists SIP over UDP, TCP and TLS as supported, DTMF under RFC 2833 and RFC 4733 supported, cold transfer by REFER and warm transfer supported — and SIP Registration (REGISTER), SIPREC and video over SIP as Not Supported. If your carrier requires a registering endpoint, that single row ends the evaluation regardless of cost. Taking the intermediary out entirely is a separate manoeuvre with its own failure set, which we set out gate by gate in running a direct SIP AI voice agent with no Twilio in the path.
Who each option is wrong for
A managed platform is wrong for you if you are sustainably above roughly 100,000 minutes a month and already employ the realtime-audio engineers, because at 200,000 minutes the framework side of our table is $8,671 a month cheaper than Bland on its Build plan, and still $761 cheaper than Vapi at its cheapest model set, even carrying the full engineering load; if you need SIP behaviour the platform does not expose; or if the per-minute rate is bundled and you have measured a model set materially cheaper than the bundle.
A framework is wrong for you if the honest answer to “who is on call” is nobody; if you are below about 38,000 minutes a month, where no row in our table that counts any engineering at all makes it pay; or if a fixed compliance cost, a call-rate ceiling or a certification you do not have is the real constraint.
And the honest case for staying managed permanently. A large number of teams will never cross 50,000 minutes a month, and for them the managed platform is not a stepping stone — it is simply the right answer, for as long as they run outbound. The rule of thumb frames the platform as a phase to grow out of. On this arithmetic it is a destination for most of the teams reading it, and the migration cost is real in both directions: what actually survives a move is a shorter list than most buyers assume, which we itemise object by object in what transfers when you switch AI voice platforms.
Zian AI sits on the other side of that boundary again: a commercial platform in partnership-application beta, with SmartReach AI™ orchestrating message, channel and timing across phone, SMS, email and WhatsApp, PrecisionPitch AI™ running continuous split-testing, 30+ languages, CRM integrations for HubSpot, Salesforce, HighLevel and Zapier, and private model deployment on customer infrastructure. The prior question — whether to own any of this — is argued separately in our build versus buy analysis for AI agents.
Frequently asked questions
Is the 10,000 minutes a month rule wrong?
On our arithmetic it reproduces only as an infrastructure-only comparison. Counting published infrastructure prices and nothing else, the framework is cheaper from about 360 minutes a month on the LiveKit Ship plan and from about 8,100 a month on Scale. Counting an amortised build plus evaluation maintenance and on-call at 5 engineer-days a month, the crossover lands between 90,000 and 183,000 minutes a month depending on which platform and which model set you compare.
How much does Vapi charge for the platform itself?
The Vapi pricing page states that usage is a hosting fee of $0.05 per minute plus model costs, and that each model component is billed at the listed price of the provider with no markup. Read 18 September 2026. A success package is priced separately, from $29 a month, and additional concurrent call lines are $10 per line per month.
Does LiveKit still need Twilio in the middle?
Not as a media bridge. The LiveKit telephony documentation states that LiveKit telephony handles inbound and outbound calls and supports integration with third-party SIP providers, and LiveKit Cloud prices third-party SIP minutes at $0.003 per minute on the Scale plan. A trunk provider is still required. Note that the same feature matrix lists SIP Registration as not supported, as at 18 September 2026.
Which per-minute rates can be compared directly?
Only the orchestration layer. Vapi at $0.05 per minute, the Retell Voice Infra line at $0.055 per minute and Pipecat Cloud agent-1x at $0.01 per minute all exclude the language model, speech-to-text and text-to-speech. The Bland AI rate of $0.12 per minute on the Build plan includes them, so comparing that headline against the others overstates the difference.
What is the biggest hidden cost on the framework side?
Not infrastructure. At 50,000 minutes a month our framework infrastructure line is $1,680 while the assumed engineering load is $8,458, so engineering is roughly five times the running cost. The evaluation suite is the part teams underestimate most, because every model or plugin migration invalidates some of it.
Does the calculation change for inbound only?
Yes, and it moves toward building. An inbound-only agent skips answering machine detection, outbound trunking and most consent machinery, which takes our assumed build from 83 engineer-days to about 22 for a single inbound line in one language. That alone moves the build-only crossover down by roughly a factor of four, and adding lines or languages adds some of it back.
Working out which side of the line you are on
If your loaded ops number is higher than half an engineer-day a month and your volume is under 50,000 minutes, the arithmetic on this page says stay managed and spend the engineering somewhere it compounds. Apply For Partnership and we will talk about the outcome you want rather than the transport you would otherwise have to run.