A8gent

Vapi Pricing and Review (2026)

Developer platform for building voice AI agents.

Reviewed by A8gent
Contents
  1. Verdict
  2. What is Vapi?
  3. Best for, and not for
  4. Strengths and limits
  5. Pricing
  6. What people actually build with Vapi
  7. What Vapi actually costs at small-business volume
  8. The gotcha: compliance costs more than the calls
  9. What breaks in production, and how you find out
  10. What it connects to that you already run
  11. Vapi alternatives: what to pick instead, and when
  12. Migration and lock-in: can you get your work out?
  13. FAQ
VerdictIs Vapi worth it?

Vapi is the developer's choice in voice AI and its pricing is honest about what that means. As of August 2026 it charges $0.05 a minute for its own hosting plus $0.005 per SMS or chat message, and passes through speech-to-text, model and text-to-speech costs at cost - or at nothing at all if you bring your own provider keys. The Build plan is pay-as-you-go with 60-plus minutes included and 10 concurrent lines, with extra lines at $10 each per month. On our 1,500-minute reference workload that is $75 of Vapi plus whatever your model and telephony stack costs, typically landing the all-in figure somewhere between $120 and $200 a month.

What is Vapi?

Vapi is a developer-focused platform for building, testing, and deploying voice AI agents over the phone and the web. It orchestrates the full real-time voice loop, letting you plug in speech-to-text, a large language model, and text-to-speech from providers you choose, while it handles turn-taking, interruptions, and low-latency streaming. You can build inbound and outbound agents, buy or bring phone numbers, and wire in tools and webhooks so the agent can take actions during a call.

Best for

Developers who want fine-grained control over the voice stack and are comfortable working with an API and their own backend.

Not for

A small business without a developer. Vapi hands you the components and expects you to assemble them, and the $0.05 headline is only the platform slice - the total is your responsibility to work out.

Strengths

  • Genuinely low platform fee at $0.05 a minute, with everything else passed through at cost rather than marked up
  • Bring your own provider keys and Vapi's share of the model and voice cost drops to zero
  • Full control over which transcription, model and voice provider you use, so you can tune cost against quality per component
  • Pay-as-you-go with no minimum commitment on the Build plan, so a pilot costs almost nothing
  • 10 concurrent calls included, with additional lines at a clear $10 each per month
  • Compliance available as explicit priced add-ons rather than a sales negotiation
  • Strong developer documentation and an API-first design, which makes deep integration practical

Limitations

  • Unbundled pricing means the headline rate is not the bill, and forecasting requires modelling four providers
  • Clearly a developer product: assembling a reliable voice agent takes real engineering
  • HIPAA at $2,000 a month and ZDR at $1,000 a month are substantial for a small business needing either
  • Latency depends on the provider combination you choose, so a bad stack produces an agent that feels sluggish
  • Concurrency beyond 10 lines is an additional cost that is easy to forget until a busy morning
  • Telephony is your problem, and phone number provisioning and carrier quirks consume more time than expected

Vapi pricing

Vapi unbundles: you pay $0.05 a minute for Vapi's own hosting and orchestration and then pay each underlying provider separately for transcription, language model and speech synthesis, at cost - or at zero to Vapi if you supply your own API keys. SMS and chat messages are $0.005 each. The Build plan is pay-as-you-go with 60-plus minutes included and 10 concurrent calls, with additional lines at $10 per line per month. The Scale plan is an annual contract with a fixed platform fee, committed volume, volume-based per-minute pricing and enterprise features including SOC 2, HIPAA, PCI, SSO and RBAC. HIPAA is a $2,000 per month add-on and Zero Data Retention $1,000 per month on either plan.

Vapi pricing tiers
PlanPriceWhat you get
Vapi hosting$0.05/minPlus $0.005 per SMS or chat message. Model, STT and TTS providers billed separately at cost, or $0 to Vapi with your own keys (as of Aug 2026)
Build (pay-as-you-go)Usage60+ minutes included, 10 concurrent calls, additional lines $10 per line/month, email and Discord support
ScaleAnnual contractFixed platform fee with committed volume, volume-based per-minute rates, SOC 2, HIPAA, PCI, SSO, RBAC, dedicated account team
HIPAA add-on$2,000/moAvailable on both plans
Zero Data Retention$1,000/moAvailable on both plans
Worked example~$75 + providers1,500 minutes a month at $0.05 = $75 to Vapi, plus your STT, LLM and TTS provider costs and telephony

Pricing reflects public plans as of August 6, 2026 and can change. Check Vapi for the latest.

What people actually build with Vapi

Vapi is a set of components rather than a finished product, so what gets built on it is whatever a developer was willing to wire together. In practice that converges on a short list of workflows, and they are worth describing concretely because the generic pitch for a voice agent hides how narrow and how valuable the good ones are.

After-hours emergency triage for an HVAC or plumbing company is the highest-value build in the category and it is not close. An inbound call rings out past six, the agent answers, and its first job is to sort emergency from non-emergency: is water actively escaping, is there no heat overnight, is anyone unsafe. On a yes it fires a tool call that pages the on-call technician's mobile with the caller's number and their own description, and escalates to a second contact if nobody acknowledges inside a few minutes. On a no it offers the next dispatch slot and writes a lead row into the field service system, so the job exists in the morning rather than in a voicemail box.

Missed-call text-back with a live callback is the dental practice version. A webhook fires when the practice line rings out, the agent dials back within a minute, calls a function against the practice management system to see whether the number matches an existing patient, and offers the next two open hygiene slots. The booking writes back, which matters more than it sounds: a booking that only exists in the call transcript is a double-booking waiting to happen.

Veterinary refill and question triage is the build that shows what guardrails are for. The agent captures pet name, medication and preferred pharmacy into a structured payload and hands it to an existing n8n or Make.com workflow that creates the task in the practice information management system. Any clinical question - is this limp serious, should I be worried about the vomiting - has to trip an escalation and transfer to a person, and that path deserves harder testing than the happy path.

Two more that recur: a restaurant reservation and waitlist line that takes party size and time, checks availability and confirms by text at $0.005 a message; and an outbound recall sweep for a medical clinic, where a nightly job pulls patients overdue for review and the agent calls in an afternoon window to confirm, reschedule or mark do-not-contact.

What Vapi is genuinely bad at

Being a product. There is no drag-and-drop path from a phone call to your booking system, so a business without a developer or an existing automation layer does not get a working agent out of Vapi at any price. It is also poor at anything requiring discretion: a caller disputing a bill, a bereaved client, a complaint that needs judgement. Those calls should be routed to a person on the first signal, not handled well.

Bottom line: The workflows that pay on Vapi are the ones where the agent gathers structured facts and hands them to a system you already run, not the ones where it is asked to exercise judgement.

What Vapi actually costs at small-business volume

Vapi's meter is the cleanest in voice AI because it only charges for its own layer. As of August 2026 that is $0.05 a minute for Vapi's hosting and orchestration, plus $0.005 per SMS or chat message. Speech-to-text, the language model and text-to-speech are passed through at cost - and drop to zero from Vapi's side entirely if you bring your own provider API keys. Telephony is separate again.

On our reference voice workload of 1,500 minutes a month - roughly a busy clinic reception or an after-hours line for a trades business - Vapi's own share is $75. That is the honest, verifiable figure. What it is not is your bill. Add transcription, a language model on every turn, voice synthesis and a phone number with per-minute carrier charges, and the all-in number is typically two to three times the Vapi line. We are deliberately not quoting per-minute rates for those providers here, because they are other companies' prices that move independently and we have not verified them on this page - price each one from its own current pricing page and add them up.

The structural point that makes Vapi cheap is unbundling. Because every component is separately chosen, you can tune cost against quality per component: a budget transcription provider, a small fast model for routing turns and a premium voice only where the caller actually hears it. Platforms that bundle everything into one per-minute rate remove that lever, and removing it is how a bundled $0.15 a minute quietly beats an unbundled $0.09 for some workloads and loses badly on others.

Concurrency is the other line item. Ten concurrent call slots are included, with additional reserved lines at $10 each per month, adjustable immediately from the billing settings. For a single-location business that never sees more than a handful of simultaneous calls, ten is plenty and the cost is zero. For anything with a morning rush, buy the lines before the rush rather than after.

Vapi all-in stack, 1,000 inbound calls a month averaging 4 minutes (4,000 minutes). Rates as published August 2026.
LayerRateBundled or separateAt 4,000 minutes
Vapi platform and orchestration$0.05/minBilled by Vapi$200.00
Speech-to-textProvider's own rateSeparate; $0 to Vapi with your own keyNot published by Vapi
Language modelProvider's own rateSeparate; $0 to Vapi with your own keyNot published by Vapi
Text-to-speechProvider's own rateSeparate; $0 to Vapi with your own keyNot published by Vapi
Telephony, US local inbound$0.0085/min plus $1.15/mo number (Twilio)Separate, billed by your carrier$35.15
Concurrency10 lines included, $10 per extra line/moBilled by Vapi$0 at 10 lines
SMS or chat messages$0.005 eachBilled by Vapi$5 per 1,000 confirmations
HIPAA add-on$2,000/moBilled by Vapi, either plan$2,000 if you need it
Verifiable subtotal--$235.15 plus three AI components

The all-in stack at 1,000 calls a month

Step the workload up to a busier business - 1,000 inbound calls averaging four minutes, so 4,000 minutes - and the shape of a Vapi bill becomes obvious. Vapi's own share is $200. Inbound US telephony on Twilio's published pay-as-you-go rate of $0.0085 a minute adds $34, plus $1.15 for the local number. That is $235.15 a month that we can verify from published pages, and it is the smaller half of the bill for most configurations, because three AI components still sit outside it.

Vapi does not publish rates for those three, and we will not invent them. The nearest published anchor is Retell AI's component table, which prices text-to-speech at $0.015 a minute and language models from $0.003 to $0.16 depending on which you pick. Applied as a sanity check rather than as Vapi's price, that suggests another $0.02 to $0.20 a minute on top - $80 to $800 at this volume. The spread is the point: on Vapi, the model and voice you choose matter far more to your bill than the platform does.

Bottom line: At 1,000 calls a month averaging four minutes, the part of a Vapi bill you can verify from published pages is $235 - and the three components nobody publishes can add anywhere from $80 to $800 on top.

The gotcha: compliance costs more than the calls

This is the number that should change decisions and almost nobody sees it coming. As of August 2026 Vapi lists HIPAA compliance as a $2,000 per month add-on and Zero Data Retention as $1,000 per month, on either plan. Set that against our 1,500-minute workload, where Vapi's usage charge is $75. The compliance add-on is roughly twenty-six times the cost of the actual calls.

For a dental practice, a physiotherapy clinic or a medical answering service - exactly the businesses for whom a voice agent is most compelling - that reframes the entire evaluation. Vapi is not a $75-a-month product for you; it is a $2,075-a-month product, and it needs to be compared against alternatives on that basis. Whether you genuinely need the HIPAA add-on is a question for your own advisers rather than for us, but find the answer before you build, because discovering it afterwards invalidates the business case rather than adjusting it.

Worth knowing how the competition prices the same thing, because it is the one place Vapi looks expensive rather than cheap. Retell AI and Bland AI both put a signed BAA on their Enterprise tier, which means a sales conversation rather than a published number. Phonely puts BAA signature on Enterprise too. Vapi is the only one of the four that will sell you compliance at a price you can read on a page, and $2,000 a month is simultaneously the highest published figure and the only published figure.

The second, smaller gotcha is the shape of the bill itself. Because model, transcription and voice costs pass through at cost, a change in your agent's behaviour changes your bill without you touching a pricing page. Switch to a more capable model for better handling of accents, or a premium voice because the cheap one sounded robotic, and the per-minute total moves. Bringing your own provider keys is the fix worth doing early: it makes Vapi's share zero on those components and puts the model spend in an account where you can see it, cap it and alert on it.

Bottom line: Vapi is the only major voice platform that publishes a compliance price, and at $2,000 a month for HIPAA it is roughly ten times the cost of 4,000 minutes of calls.

What breaks in production, and how you find out

Concurrency is the first wall, and Vapi's documentation is unusually specific about what happens at it. Every account includes 10 concurrent call slots by default; each call occupies one slot like a physical phone line. When all slots are busy, new outbound dials and inbound connections wait for a slot to free up rather than failing outright. Crucially, the API tells you: responses carry a `remainingConcurrentCalls` field, and `concurrencyBlocked` becomes true when a call cannot start. Those two fields are the difference between knowing you are capacity-constrained and guessing, and any serious deployment should be logging and alerting on them from day one. Reserved concurrency can be raised immediately from Settings then Billing.

Latency is the failure that ruins the experience without ever registering as an error. Vapi's own engineering blog states that conversational flow breaks when latency exceeds roughly 1,200 milliseconds - that is the vendor's own figure and we cite it as such, but it matches the intuition anyone gets from a first test call. Past that threshold the caller starts talking over the agent, the agent interrupts, and the conversation degrades into the thing people hate about automated phones. Every component you add - a bigger model, a fancier voice, an extra tool lookup mid-turn - spends part of that budget. Measure end-to-end latency on real calls, not in the playground.

The third pattern is architectural, and it is a common observation in the developer community rather than a documented limit: deployments fall over at scale because a single webhook endpoint becomes a bottleneck during traffic surges, and because teams treat a voice platform like a single-server application instead of queueing work. Vapi documents call queue management for exactly this reason. If your calls trigger downstream work - a CRM write, a booking, a notification - put a queue between the call and that work rather than doing it inline.

Finally, retention. On the Build plan, call history is kept 14 days and chat history 30 days as of August 2026, with custom terms on Scale. Two weeks is a short window for investigating a customer complaint, and it is not an audit trail. Push transcripts and call outcomes into your own CRM or database as they happen if you need them to exist next quarter.

Sizing concurrency for a real phone queue

Ten included lines sounds generous until you translate it. A business taking 1,000 calls a month is averaging around 45 on a working day, but those calls are not evenly spread - a clinic gets a spike the moment the doors open and another at lunchtime, and a trades business gets one the morning after a cold snap. Twenty simultaneous calls is a realistic peak for a business at that volume, and on Vapi that is ten extra reserved lines at $10 each, so $100 a month on top of the $200 of minutes. That is a 50 percent uplift on the platform line for capacity you use for twenty minutes a day, and it is the single most commonly missed item in a voice AI budget.

The good news is that Vapi's behaviour at the ceiling is queueing rather than rejection, so the failure is a wait rather than a busy signal. Alert on `concurrencyBlocked` and buy lines from the data rather than from a guess.

Bottom line: Vapi queues rather than rejects at the concurrency ceiling, and the `remainingConcurrentCalls` and `concurrencyBlocked` fields are the only warning you will get before callers start waiting.

What it connects to that you already run

Vapi's integration philosophy is the same as its pricing philosophy: it gives you the primitives and expects you to wire them. The provider side is genuinely open - you choose your transcription, language model and voice vendors and can bring your own keys for each, which is what makes the per-component cost tuning possible. Telephony is likewise bring-your-own or buy-through, so an existing business number can usually be brought to the platform rather than replaced.

For business systems, the connection point is tool calling and webhooks rather than a connector catalogue. Your agent calls a function mid-conversation to check availability, look up a customer or create a booking, and that function is an endpoint you host or an automation you have already built. In practice that means Vapi pairs naturally with whatever automation platform you already run - the voice agent handles the conversation and hands structured data to an n8n, Make or Zapier workflow that does the actual work against your CRM, calendar or practice-management system.

That is a strength and a prerequisite in equal measure. If you have a developer or an existing automation layer, Vapi slots into it beautifully and you are not limited to whatever integrations a vendor chose to build. If you do not, there is no drag-and-drop path from a phone call to your booking system, and that gap is the actual reason Vapi projects stall for small businesses.

Bottom line: Vapi connects to your business through function calls and webhooks, which means it pairs with the automation platform you already run rather than replacing it.

Vapi alternatives: what to pick instead, and when

The honest framing for Vapi alternatives is not which platform is cheapest per minute - it is how much of the stack you want to own. Vapi sells you the orchestration and leaves speech-to-text, the language model, text-to-speech and telephony as your problem, priced by other companies. Every alternative below moves that line somewhere else, and where it sits is the whole decision.

Pick Retell AI if you want Vapi's component-level control without doing the arithmetic yourself. Retell publishes every rate - voice infrastructure at $0.055 a minute, text-to-speech from $0.015, models from $0.003 to $0.16, US telephony at $0.015 - so you can build an accurate forecast before you write a line of code. It also includes 20 concurrent calls against Vapi's 10, which for a business with a real phone queue is worth more than a cent or two on the rate.

Pick Bland AI if nobody in the business writes code. One bundled per-minute rate covering the model, transcription and premium voices, no platform fee on the entry plan, and a working agent in days rather than weeks. It costs more per minute at this volume and you lose the ability to tune components, which is exactly the trade a business without a developer should take.

Pick ElevenLabs Agents when the phone call is part of your brand and a caller hearing a natural voice matters more than a few cents a minute. Pick Phonely if your call pattern is spiky rather than steady, because unlimited concurrency on every paid plan removes the constraint Vapi charges $10 a line to relieve. Pick Synthflow only if you are a contact centre - its published floor is $30,000 a year, which is a different market rather than a different plan. And if you are weighing whether to run any of this yourself, our page on what automation costs sets the build against the licence.

Vapi alternatives at 1,000 inbound calls a month averaging 4 minutes (4,000 minutes), on published August 2026 rates.
PlatformBilling unitAt 4,000 minutesWho it suitsDoes better than Vapi
Vapi$0.05/min platform, everything else at provider cost$235 verifiable, plus 3 AI componentsTeams with a developer and an existing automation layerCost transparency and free choice of every provider
Retell AIItemised per-minute components$354 lean, $442 balanced, $1,082 premiumTeams wanting the same control with the pricing already publishedForecastable before you build; 20 concurrent calls included
Bland AIOne bundled per-minute rate$560 on Start, no platform feeBusinesses with no engineering capacityNothing to assemble; live in days
ElevenLabs AgentsSubscription minutes at a flat $0.08$320 plus model and telephonyClinics and premium services where the voice is the first impressionVoice quality, audibly
PhonelySubscription minutes with high overage~$1,125 on Pro at $0.30 overageSpiky, bursty call patternsUnlimited concurrency on every paid plan
SynthflowEnterprise annual contract$2,500/mo floor, no per-minute rate publishedContact centres with existing telephonyScoped launch support and integration services

Bottom line: Choose Vapi over its alternatives only if someone will own the assembly; if nobody writes code, Bland AI's higher per-minute rate is cheaper than the engineering time Vapi requires.

Migration and lock-in: can you get your work out?

Vapi has one of the better lock-in positions in voice, and the reason is unbundling. Because you choose the transcription, model and voice providers yourself and can hold your own API keys with them, the components that do the actual work are not Vapi's to withhold. Your relationship with your model provider survives you leaving Vapi. That is a meaningfully different position from a bundled platform where every layer is the vendor's.

The phone number is the asset that genuinely binds you, as with every voice platform. Bringing your own number from a carrier you already control, rather than provisioning one through the platform, converts your biggest switching cost into a configuration change. If you must provision through the vendor, establish the porting terms before you print the number on anything.

What you would rebuild on a move is Vapi's own layer: the assistant configuration, the tool definitions, the conversation design and the webhook plumbing. That is real work, but it is configuration and prompts rather than proprietary logic, and it is readable. Keep the conversation design - what the agent says, what it must never say, how it escalates to a human, what it does when it does not understand - in a document outside the platform. That document is the thing worth months of refinement, and it ports to any vendor.

The commercial side is friendly: the Build plan is pay-as-you-go with no minimum commitment, so leaving costs you nothing but the rebuild. The Scale plan is an annual contract, so if you sign one, the renewal date becomes your exit window - worth weighing against the volume discount it buys.

Bottom line: Because your model, voice and carrier accounts are your own, leaving Vapi means replacing orchestration rather than rebuilding a stack - the mildest lock-in position in voice AI.

Vapi FAQ

What does Vapi actually cost per month?

Take the reference workload of 500 calls a month averaging three minutes, so 1,500 minutes. Vapi's own charge as of August 2026 is $0.05 a minute, so $75. On top of that you pay transcription, model and speech synthesis at provider cost, and telephony separately - which in practice adds somewhere around $0.03 to $0.08 a minute depending on the voice and model you pick. All in, expect roughly $120 to $200 a month for that volume, with the spread driven almost entirely by voice quality.

What does bring your own keys mean for cost?

If you supply your own API keys for the transcription, model and speech providers, those costs bill directly to your accounts with them and Vapi's share of them is zero - you pay Vapi only its $0.05 a minute hosting fee. That is genuinely useful for two reasons: you get real visibility into what each component costs, and you can negotiate or optimise those relationships independently. It also means a model price change hits your bill directly rather than being absorbed and repriced.

What breaks with voice agents in production?

Latency and interruption handling, before anything else. A voice agent that pauses for a second before answering feels broken to a caller in a way a slow chatbot never does, and the pause is usually the sum of four providers' response times. Second is background noise and accents degrading transcription, which then propagates into wrong answers. Third is the agent failing to recognise when a caller wants a human. Test with real callers on real phones, not with your own voice in a quiet office.

Who should not use Vapi?

A small business without engineering capacity. Vapi is a set of well-designed components with an excellent price, and assembling them into something a customer will tolerate is real work - choosing providers, tuning latency, handling interruptions, wiring telephony. If you want a voice agent live next week and nobody in the building writes code, a bundled provider costs more per minute and vastly less in total. Be honest about which situation you are in.

What does compliance cost?

It is priced explicitly, which is refreshing: HIPAA is a $2,000 per month add-on and Zero Data Retention is $1,000 per month, available on both plans as of August 2026. For a small clinic or practice, $2,000 a month is a serious number that may well exceed the value of the automation - so if you are in a regulated sector, price the compliance add-on first and decide whether the project still makes sense before doing any build work.

How locked in are we?

Comparatively little, which is a real advantage of the unbundled model. Your model, voice and transcription providers are your own accounts, your phone numbers are with your carrier, and your prompts and call scripts are portable text. What Vapi provides is orchestration, so leaving means replacing that layer rather than rebuilding the stack. That is a meaningfully better position than a bundled provider where the voice, the model and the numbers are all theirs.

How many concurrent calls do we need?

More than you think, and it is the limit most people hit first. Concurrency is about simultaneous calls, not monthly volume: a business taking 500 calls a month can easily see six at once on a Monday morning. Vapi includes 10 concurrent lines on the Build plan with extras at $10 per line per month, which is generous. Model your peak hour rather than your monthly total, because callers who hit a busy signal do not call back.

Vapi or a bundled voice provider?

Vapi if you have engineering capacity and want cost transparency and control over every component. A bundled provider if you want a working agent quickly and are willing to pay a higher per-minute rate for someone else to have solved the integration. The per-minute gap is real but usually smaller than the cost of the engineering time, so for most small businesses without a developer the bundled option is genuinely cheaper overall.

Can you build a phone agent for us?

Yes, and missed-call handling and appointment booking are the two we are asked for most - they are also the two where the value is easiest to prove, because you can count the calls you currently miss. We scope it on a free automation audit, pick the platform on your volume, compliance needs and whether anyone will maintain it, and hand it over with the prompts, the number and documentation you own, plus a recurring report of what it handled and what it escalated.

Looking at alternatives? Vapi is the right choice when you have engineering capacity and want control over every layer of the voice stack, and its pricing transparency is genuinely better than the bundled alternatives - you can see exactly what each component costs. The flip side is that you own the assembly and the reliability. On lock-in it is in a good position: because the model, voice and telephony providers are yours, moving off Vapi means replacing orchestration rather than rebuilding your whole stack. If nobody in the business writes code, a bundled provider will get you live in days rather than weeks. See the Autonoly review.