An AI voice agent platform is software that runs a voice agent on your phone lines: it answers or places the call, talks with the person in their own words, completes the task, and transfers what it cannot finish. Six of the better known ones publish prices, and no two price the same call the same way, so the figure on one pricing page will not line up against the figure on the next.
That is the whole problem with shortlisting one. A per-minute rate on one site covers the language model, the transcription and the voice. On another it covers the orchestration layer and nothing else. A third quotes a range four times wide at the top end than the bottom, and which end you land on depends on options you pick later. Below are the published numbers as of 2026-10-18, what each one actually includes, and the arithmetic that turns any of them into a cost per contact for your own call mix.
Telvana is one of the six, and it is ours. Every figure here comes from each platform's own current public pricing page, linked where it is used.
Three ways these platforms are sold
The pricing shape follows from the purchase, not from the vendor. Most of these companies will sell you more than one of the three, so decide which one you are buying before you compare any rates.
| How it is sold | What you get | Who owns the agent afterward | What the price looks like |
|---|---|---|---|
| Developer platform | An API, a dashboard and the building blocks. You assemble the agent | Your engineers | Per minute, often split across components |
| Bundled agent platform | A builder plus an API, with limits set by plan tier | Your operations or RevOps team | One per-minute rate, or minutes bundled into a tier |
| Scoped deployment | The vendor's team builds and tests the agent with you | The vendor, with your approval | A contract, or a monthly price sized to volume plus a build fee |
What sorts you into a row is a staffing question. Who writes down the procedure the agent follows, and who fixes it at 7am on a Monday when the transfer rule turns out to be wrong? If the answer is nobody on your payroll, a developer platform will be cheap on the quote and expensive in practice.
For what these agents can and cannot do on a call, separate from who sells them, see AI call center: what AI agents handle and how to add them. If the decision in front of you is whether to automate a call type or send it to people at all, call center outsourcing vs AI agents has a scorecard for that one, and the rest of our writing on this sits under choosing a platform.
What AI voice agent platforms charge, side by side
Published prices as of 2026-10-18. Read the right-hand column before the rate, because the rates do not cover the same things.
| Platform | Entry price | Per-minute rate | What the rate covers |
|---|---|---|---|
| Vapi | $5 in credits, no platform fee | $0.05 per minute hosting, plus components | Orchestration only. Transcription, model and voice are billed separately at provider rates |
| Retell AI | $10 in credits, no platform fee | $0.07 to $0.31 per minute | Varies with the model, voice, telephony and add-ons you select |
| Bland AI | No platform fee on Start | $0.14 per minute, or $0.12 with a $299 monthly fee | Model, speech to text and text to speech in one rate. Telephony billed separately |
| ElevenAgents | Free tier with 15 minutes | Minutes bundled into the plan | 1,238 minutes at $99, 3,738 at $299, 12,375 at $990 per month |
| Synthflow | Contracts start at $30,000 a year | Not published | Scoped around volume, concurrency, telephony, integrations, security and launch support |
| Telvana | From $399 a month | Not published | Monthly price sized to your call volume, plus a one-time setup fee |
Before any of the detail, there are two things in that table worth sitting with.
The first is that two of the six publish a floor and no rate, and Telvana is one of them. If you need a published per-minute number to model this at all, that rules us out at the shortlist stage, and it should. We quote the monthly and the setup fee on a demo once we know your volume and your call types.
The second is that Vapi's own calculator is the clearest illustration of the problem. Enter 1,000 minutes and it returns $82 to $129 a month, built from four separate lines: $50 of Vapi hosting, $10 of Deepgram transcription, $8 to $45 of OpenAI, and $15 to $24 of ElevenLabs voice (Vapi pricing). The headline $0.05 is under a third of the real total at the low end, and under half of it at the high end.
Bland says the same thing about its competitors, in its own words: their advertised per-minute rate "is the platform fee only," and adding typical provider rates "brings most production stacks to roughly $0.13 to $0.30 per minute on Vapi and $0.11 to $0.25 per minute on Retell" (Bland AI pricing FAQ). Read that as a competitor's framing, then check it against the component rates Vapi publishes, which is where the $0.13 to $0.30 comes from.
Why the per-minute rate is not comparable
A per-minute rate is comparable only when the two rates cover the same components. Right now they do not, and the gap between the cheapest and dearest reading of one platform can be wider than the gap between two different platforms.
Take Retell's published $0.07 to $0.31 per minute. At 20,000 minutes a month that is $1,400 at the bottom and $6,200 at the top, from the same price list. Which number you get depends on choices its cost estimator exposes: which language model, which voice provider, whose telephony, and whether you add the knowledge base, denoising, safety guardrails, PII removal or AI quality assurance (Retell AI pricing). None of those is an exotic requirement for a contact center.
ElevenAgents shows the opposite pattern. Divide each plan price by its included minutes and the answer is the same at every business tier: $99 for 1,238 minutes, $299 for 3,738, and $990 for 12,375 all work out to $0.08 per included minute (ElevenAgents pricing). Climbing that ladder does not buy a better minute. It buys concurrency and seats, which is a different purchase, and one worth making deliberately rather than by accident.
What concurrency costs on each platform
Concurrency is the number of calls the platform will run at the same time, and for a contact center it is usually the constraint that binds first. Minutes accumulate over a month. Concurrency has to be there in the busiest half hour or callers get nothing.
Published caps and prices as of 2026-10-18:
| Platform | Included concurrency | More of it |
|---|---|---|
| Retell AI | 20 free | $8.00 per concurrency per month |
| Vapi | 4 on usage only, 10 on Core at $29, 30 on Pro | $10 per line per month |
| Bland AI | 10 on Start, 50 on Build at $299 | Enterprise, sized to volume |
| ElevenAgents | 4 free, 20 at $99, 30 at $299, 40 at $990 | Enterprise, elevated limits |
| Synthflow | Planned in the contract | Part of the scope |
| Telvana | Not published | The monthly is sized to your call volume |
Now price one requirement across them. A center that needs 40 simultaneous AI calls pays $160 a month in concurrency on Retell, on top of usage. The same 40 lines mean Vapi Pro at a $999 monthly minimum plus ten extra lines at $10 each, so $1,099 a month before a single minute. On Bland it is covered inside the $299 Build fee, because Build includes 50. On ElevenAgents it lands exactly on the $990 Business tier. The requirement is identical in all four cases and the fixed cost of meeting it varies by about seven times.
How many concurrent calls does your center need?
Multiply your busiest hour's calls by your average handle time in minutes, then divide by 60. That is the average number of calls in progress during that hour, and it is the floor for any concurrency cap you buy.
average calls in progress = (peak hour calls x AHT in minutes) / 60
600 calls x 4 minutes / 60 = 40
200 calls x 4 minutes / 60 = 13
120 calls x 5 minutes / 60 = 10
This is the offered load, the quantity the call center literature writes as arrival rate times average service time (Gans, Koole and Mandelbaum, Telephone Call Centers: Tutorial, Review, and Research Prospects, Manufacturing and Service Operations Management 5, 2003). It is an average, and arrivals inside the hour are not evenly spaced, so a cap set at exactly that number will turn callers away during the spikes inside it. Size above it, and use the staffing calculator if you want the proper sizing rather than the floor.
If your handle time is measured four different ways across your reports, settle that first. Average handle time shows how one half hour produces several defensible AHT numbers, and this calculation inherits whichever one you feed it.
A cost per contact worksheet
Four lines turn any quote, in any pricing shape, into a number you can compare and put in a business case. Fill them in per call type rather than for the whole center, because handle time and volume both change from one call type to the next.
| Line | What to enter | Example |
|---|---|---|
| A. Contacts a month | Volume of the one call type you are automating | 5,000 |
| B. Average handle time | Minutes, for that call type | 4.0 |
| C. All-in per-minute rate | Every component the platform bills, not the headline | varies |
| D. Fixed monthly | Platform fee, concurrency add-ons, compliance add-ons, seats | varies |
Cost per contact is then (B x C) + (D / A).
Run the four lines across the published rates at 5,000 contacts and a 4 minute handle time, which is 20,000 minutes a month:
| Platform and plan | Usage | Fixed | Total a month | Per contact |
|---|---|---|---|---|
| Bland Build, $0.12 per minute | $2,400 | $299 | $2,699 | $0.54 |
| Bland Start, $0.14 per minute | $2,800 | $0 | $2,800 | $0.56 |
| Retell, $0.07 to $0.31 per minute | $1,400 to $6,200 | $0 | $1,400 to $6,200 | $0.28 to $1.24 |
| Vapi components plus Twilio inbound, Core | $1,796 to $2,738 | $29 | $1,825 to $2,767 | $0.37 to $0.55 |
| Vapi components plus Twilio inbound, Pro | $1,796 to $2,738 | $999 | $2,795 to $3,737 | $0.56 to $0.75 |
| Synthflow contract floor, before usage scoping | not published | $2,500 | $2,500 and up | $0.50 and up |
Telephony is in the Vapi rows at the $0.008 per inbound minute Vapi lists for Twilio, and excluded from the Bland rows, because Bland bills it separately or lets you bring your own carrier. Add your own carrier's rate to any row that does not include one.
The ranking is the least interesting thing in that table. What matters is that the spread inside Retell's own published range, 28 cents to $1.24, is wider than the spread between every other platform on the list, so a shortlist built on headline rates is sorting noise rather than cost. The cheapest cell also belongs to the platform with the widest uncertainty, and procurement spreadsheets tend to reward exactly that.
When a platform fee pays for itself, and when a cap stops you first
Divide the monthly fee by the per-minute saving it buys. Bland's $299 Build fee takes the rate from $0.14 to $0.12, so it pays for itself at $299 / $0.02, which is 14,950 minutes a month. At a 4 minute handle time that is about 3,737 calls, or 125 calls a day.
Except you will never get there on Start, because Start caps at 100 calls a day (Bland AI pricing). At a 4 minute handle time that cap tops out at 12,000 minutes a month, which is below the 14,950 where the fee starts paying. The cap moves you up a tier before the price ever does.
That pattern repeats across the category, so it is worth checking early. Published caps on daily calls, hourly calls, concurrency, knowledge bases, voices and data retention decide which tier you are on. The per-minute rate only decides what that tier costs. Work out your tier from the caps first, then compare rates inside it.
A shortlist scorecard for a contact center
Score each candidate out of 10 on these seven, multiply by the weight, and total. The weights are a planning heuristic rather than a validated benchmark, so change them where your center's constraints differ, but change them before you score, not after.
| Criterion | Weight | What a 10 looks like |
|---|---|---|
| Concurrency at your peak | 25% | Covers your offered load with headroom, at a fixed cost you can state |
| Who owns it after launch | 20% | Matches who on your side will actually maintain the agent |
| What the quoted rate covers | 15% | Every component named, no separate provider bills to discover later |
| Transfer behavior and rules | 15% | Transfer conditions and destinations you set, tested before launch |
| Call records and outcome tagging | 10% | Every call recorded, transcribed and tagged, exportable |
| Separate workspaces per program | 10% | One isolated workspace per team or client program |
| Compliance paperwork you need | 5% | The specific agreement your legal team asks for, in writing |
The concurrency row is the one most often filled in from a hope rather than a measurement, so compute the offered load before you score anything. For a BPO the workspace row is closer to a pass or fail than a score: running several client programs out of one workspace means one client's knowledge, numbers and call records sitting next to another's.
What to check before you sign
These six are listed in the order they tend to surprise people.
- Ask for the list of separately billed components in writing: transcription, model, voice, telephony, concurrency, data retention, compliance add-ons. Vapi's HIPAA handling, for one example, is a $2,000 a month add-on (Vapi pricing).
- Check the published caps against your busiest day rather than your average one. Daily and hourly call caps bind, and they are the reason most centers end up a tier higher than they planned.
- Ask how calls will reach the agent. Any phone system that can transfer or forward a call to an outside number can send calls to Telvana, and the others range from built-in telephony to bring-your-own SIP, so confirm which applies before you plan a migration you may not need.
- Pin down the transfer. What conditions trigger one, what number does it land on, and what does the caller hear? Telvana's transfers are cold: the call moves to the number you choose under rules you set, and the person who picks up is not handed a summary.
- Read the compliance claims against what your legal team actually asked for. Bland publishes SOC 2 Type I and Type II, HIPAA eligibility with a signed BAA, GDPR and PCI DSS (Bland AI pricing). Telvana holds none of those certifications, which matters if your contract requires one.
- Write the test set yourself, whoever builds the agent. Twelve conversations that catch invented policy and dead-end loops is the set we run.
Once a shortlist survives all six, what is usually left to decide is which call type to give the agent first, and that comes out of last month's contact reasons rather than any vendor's feature list.
Frequently asked questions
What is the best AI voice agent platform?
There is no single best one, because the platforms are sold in three different shapes and the right shape depends on who will own the agent after launch. A team with engineers who want to build and maintain it gets more from a developer platform. A team without one gets more from a scoped deployment, even at a higher published floor.
How much do AI voice agent platforms cost?
Published rates as of 2026-10-18 run from $0.05 per minute of orchestration on Vapi, with transcription, model and voice billed on top, to $0.14 per minute all in on Bland, with contracts from $30,000 a year on Synthflow and monthly pricing from $399 on Telvana. At 5,000 contacts a month and a 4 minute handle time, the published rates work out between about 28 cents and $1.24 per contact.
Why do AI voice agent platforms show such different prices?
Because the rates cover different components. One platform's per-minute price is the orchestration layer alone, another's includes the language model, speech to text and text to speech, and a third bundles minutes into a plan tier. Until you list what each rate includes, the numbers are not measuring the same thing.
Do I need a developer to use an AI voice agent platform?
For a developer platform, yes, and for ongoing maintenance rather than just the build. Bundled platforms add no-code builders that an operations team can run. Scoped deployments put the build on the vendor's team, which is the model Telvana uses: we build and test the agent from your scripts, policies and transfer rules, and you approve it before it takes a live call.
How many concurrent calls do we need?
Multiply your busiest hour's calls by your average handle time in minutes and divide by 60. That gives the average number of calls in progress in that hour. Buy above it, because arrivals inside the hour are uneven, and check the figure against each platform's published concurrency cap before you compare per-minute rates.
Can we try one before committing?
On the self-serve platforms, yes, and the published free credits are real: $5 on Vapi, $10 on Retell, 15 minutes of calls on the ElevenAgents free tier. On the scoped side there is no self-serve trial, so ask instead to hear an agent handle one of your own call types during the evaluation.
See which of your calls Telvana can take
Book a demo, bring your top call types, and we will show you which ones an agent can handle, let you hear it take one, and quote the monthly and setup fee against your actual volume.