AI Voice Agent Pricing 2026: What 1,000 Minutes Costs on 5 Platforms

One workload, 1,000 minutes a month of inbound US calls on a local number, one agent, no enterprise contract. On the cheapest honest configuration I could price it at $39.80 a month. On the most expensive it works out at about $2,500 a month. Same minutes, same job, a sixty-fold spread.

Almost none of that gap comes from the product. It is about what each platform counts and who assembles the stack.

Here is the part that costs people money. On three of the five, the rate on the pricing page is a fraction of the bill, not the bill. Vapi itemises transcription, model and voice as separate lines you add up yourself. Retell publishes a range that moves with your model choice, from $0.07 to $0.31 a minute. ElevenLabs sells minute buckets, so your real per-minute cost depends on how much of the bucket you actually finish.

I read every one of these pricing pages on 4 October 2026 with a desktop browser, then costed the same workload across all five and kept the arithmetic. I did not build an agent on any of these platforms, did not place a call, and did not benchmark latency or voice quality. Where latency or quality comes up below, it is the vendor's claim about their own product and I say so.

This page covers the platforms you build on: Vapi, Retell AI, Bland, ElevenLabs Agents and Synthflow, plus the telephony layer underneath them. If you want a working phone agent this week and have no intention of writing code, the done-for-you AI receptionist services are a different market with a different price floor, and that comparison is the right page for you.

How I compared them

Comparison posts in this category tend to line up feature checkboxes. Builders rarely lose on feature parity. They lose because five pricing pages describe five different units, so nothing on them is comparable until you force them onto one workload.

So I did exactly that, and nothing more than that. Here is the whole method:

• I read the vendor's own pricing page, and only that. Every figure below came off the vendor's own live page: bland.ai/pricing, retellai.com/pricing, vapi.ai/pricing, elevenlabs.io/pricing/agents, synthflow.ai/pricing and twilio.com's US voice page, all read on 4 October 2026.
• I fixed one workload: 1,000 minutes a month of inbound US calls on a local number, one agent, no enterprise contract, cheapest honest configuration.
• I showed the arithmetic. Where a platform publishes component rates, I add them and multiply. Where it sells a bucket, I divide the bucket.
• I recorded what did not load. Synthflow's self-serve tiers are rendered in JavaScript and did not appear for me, so I am listing them as unknown and quoting no number for them.

What I did not do matters just as much. I built nothing on these platforms. I ran no calls, measured no latency, and formed no opinion of my own about how any of these voices sound. A per-minute rate is a thing you can read and check. Voice quality is not, and I am not going to pretend otherwise to make the post feel more authoritative.

What 1,000 minutes actually costs

Below is the same workload priced five ways. Read the middle column before the right-hand one, because the middle column is what the right-hand one is made of.

Prices read off each vendor's own pricing page on 4 October 2026, monthly, for 1,000 minutes of inbound US calls on a local number with one agent and no enterprise contract. Twilio's US voice page states its pricing is current as of August 2026. Your own model choice, concurrency and call mix will move these numbers.

The arithmetic, so you can check it. Vapi at the low end is $0.0095 for Deepgram transcription, plus $0.0077 for the OpenAI model, plus $0.0146 for the ElevenLabs voice, plus $0.008 for Twilio inbound. That is $0.0398 a minute, so $39.80 for the thousand. At the high end the same four lines read $0.0099, $0.0452, $0.0238 and $0.008, which is $0.0869 a minute and $86.90 for the thousand. Vapi's optional Core success package at $29 a month sits on top of either figure.

Retell is the easiest to compute because Retell did it. Their cost calculator itemises a sample build at $0.11 a minute: $0.04 for the LLM, $0.055 for Retell's voice infrastructure, $0.015 for text to speech, $0.00 for telephony. A thousand minutes is $110. Their published range of $0.07 to $0.31 a minute means the same workload lands near $70 on a small model and near $310 on a realtime one.

Bland Start is the shortest sum on the page. One thousand minutes at $0.14 is $140, and the platform fee is zero.

ElevenLabs needs a division instead of a multiplication. Pro is $99 a month for 1,238 minutes, which is $0.080 a minute if you finish the bucket. Use 250 of those minutes instead and you paid $0.396 a minute. The bucket model rewards accurate forecasting and punishes a quiet month.

Synthflow is the $2,500 row, and it is a division too: $30,000 a year over twelve months. That figure is the entry point to a scoped contract, and the scope is what sets your real bill. It belongs in a table about what you actually pay for exactly that reason.

What the advertised rate leaves out

There are two billing philosophies in this market and the whole pricing argument sits on the seam between them.

Bundled. One per-minute number covers the stack. Bland's page frames its rate this way, with telephony passed through, and states plainly: "No surprise bills." You trade component-level control for an invoice you can predict in your head.

Bring your own stack. You pick the transcriber, the model and the voice, and you pay each one. Vapi is the clearest example because it is the only platform in this set that publishes per-component rates on its pricing page. Retell sits between the two: pay as you go on one number, with the component split published so you can audit where that number comes from.

Bland makes the case against its competitors on its own FAQ, and it is worth quoting because it comes from a vendor with a stake in the answer. Bland's FAQ says: "Their advertised per-minute rate is the platform fee only. Vapi lists $0.05/min and Retell lists about $0.07/min for their voice infrastructure. You then pay separately for the LLM, speech-to-text..."

Treat that as Bland's claim about Bland's rivals, not as a verified figure. I could not find a flat $0.05 a minute platform fee on vapi.ai/pricing when I read it on 4 October 2026; what that page publishes instead is the component itemisation I used in the table above. The structural point in Bland's FAQ holds anyway, and the table proves it from the other direction: Vapi's $39.80 is only $39.80 because somebody chose four components and will own the consequences of that choice.

One more thing comparison posts in this category skip. Two vendor pages disagree about Twilio. Vapi's page lists Twilio inbound at $0.008 a minute; Twilio's own US voice page lists $0.0085 to receive on a local number. The difference is five ten-thousandths of a dollar, which is 50 cents a month at 1,000 minutes and $50 a month at 100,000. Component pricing means somebody has to care about rounding at the fourth decimal place, and that somebody is you.

1. Vapi: component pricing, lowest floor, most assembly

Vapi is the only platform here that shows you the parts list. The pricing page publishes a rate for each layer: Deepgram transcription at $0.0095 to $0.0099 a minute, the OpenAI intelligence model at $0.0077 to $0.0452 a minute, and the ElevenLabs voice at $0.0146 to $0.0238 a minute. Usage-based, with $5 in free credits, two organizations and one included Vapi phone number to get started. All as of October 2026.

That transparency is the reason the Vapi column in my table spans a range at all. The model line alone moves nearly sixfold between its ends. Pick a small model and your voice agent is the cheapest thing on this page. Pick a realtime one and you have quietly doubled the bill.

Where the money actually goes

At the low end of Vapi's own published rates, the model is about 19% of the per-minute cost and the voice is about 37%. At the high end the model becomes 52% and the voice drops to 27%. In other words, the single decision that determines your voice agent's unit economics on Vapi is which LLM you point it at, and it is a decision you make in an afternoon and live with for a year.

What the $29 success package buys

Vapi's success packages are support, not usage. Core at $29 a month covers two organizations, five included phone numbers, 10 concurrent calls, 30 days of raw data retention and email support with a two-business-day SLA. Pro is stated as 10% of your Vapi hosting fee with a $999 a month minimum, and brings 10 organizations, 10 included numbers, 30 concurrent calls and 180 days of retention. Note what that means at this workload: the $999 minimum is more than ten times the entire component bill for 1,000 minutes, so Pro is priced for volume you do not have yet.

Best for

A developer or agency that wants the lowest possible unit cost and accepts owning the stack, the model choice and the maintenance that follows.

Pricing

Usage-based on published component rates. Success packages: Core $29 a month, Pro from $999 a month stated as 10% of hosting fee. As of October 2026.

Free tier

$5 in free credits, two organizations, one included Vapi phone number.

Verdict

The cheapest answer on the page, and the one that bills you in your own time as well as in dollars. If per-component pricing appeals to you for the same reasons self-hosting does, the tradeoffs in open-source AI agent frameworks will feel familiar.

2. Retell AI: the component split, published, with nothing to sign

Retell is the one I would open first if I had a client deadline. Zero to start, $10 in free credits, full platform access, no commitment, and 20 concurrent calls included on pay as you go with more available on demand. Voice agents run $0.07 to $0.31 a minute and chat agents are quoted from $0.002 a message. Those were the figures on retellai.com/pricing on 4 October 2026.

Their page makes the no-contract point itself: "Most platforms require annual contracts before you write a single line of code. Retell starts at $0, pay only for what you use."

The cost calculator is the real feature

Retell publishes a sample build itemised to the half-cent, and that is unusual enough to be worth the click. The split: $0.04 a minute for the LLM, $0.055 for Retell's own voice infrastructure, $0.015 for text to speech, $0.00 for telephony, $0.00 for add-ons. Total $0.11 a minute, or $110 for this workload. You can take that breakdown to a client and defend every line of it.

The range is the thing to watch, and it is wide for an honest reason. At $0.07 a minute the thousand costs $70. At $0.31 it costs $310. Anyone quoting you "Retell is seven cents a minute" has picked the bottom of a published range and not told you which model produces it.

Best for

An agency or technical founder who needs something billable running this week, with a cost breakdown they can show a client.

Pricing

Pay as you go from $0. Voice agents $0.07 to $0.31 a minute, chat agents from $0.002 a message, enterprise custom. As of October 2026.

Free tier

$10 in free credits with full platform access and no commitment.

Verdict

The best risk-adjusted first move here. Start on the free credits, measure your real per-minute cost on your own model choice, then decide whether Vapi's floor is worth the assembly.

3. Bland: one bundled number, and a crossover worth knowing

Bland sells the opposite proposition to Vapi. One per-minute rate covers the stack, telephony is passed through, and the pricing page says there are no surprise bills. For a lot of teams a forecastable bill is worth more than a cheap one. I have written up where Bland and a general task platform actually diverge in Gravity vs Bland AI, which exists mostly to say they are not the same purchase.

Two self-serve tiers, as of October 2026. Start is $0.14 a minute with no platform fee, no card required, 2 credits plus an inbound number that their page calls a $15 a month value, transfer minutes at $0.05, 10 concurrent calls, 100 calls a day and 10 knowledge bases. Build is $0.12 a minute plus a $299 a month platform fee, with transfers at $0.04, 50 concurrent calls, 2,000 calls a day and 50 knowledge bases. Enterprise is a custom rate with no daily cap.

Note also what the tiers actually gate. Start's 100 calls a day is a volume ceiling as much as the price is, and 10 concurrent calls is a ceiling a small inbound spike will hit. Build's real product is headroom: 50 concurrent, 2,000 a day, 50 knowledge bases. People read those two rows as a discount and buy the wrong one.

When Build beats Start

At 14,950 minutes a month, and not a minute sooner. Build saves you $0.02 a minute against Start, so you need the $299 platform fee back out of that saving: $299 divided by $0.02 is 14,950 minutes. Below that line the fee costs more than the discount returns, and the gap is brutal at small volume. At the 1,000 minute workload in the table, Start is $140 and Build is $419. You would pay three times as much for the same thousand minutes.

If you are running 14,950 minutes a month you are doing roughly 500 minutes a day, which is where Start's 100 calls a day cap starts to bite anyway. The two thresholds land close enough together that the crossover is usually a volume decision dressed up as a pricing one.

Best for the team that wants a per-minute figure it can put in a spreadsheet and stop thinking about. Pricing as of October 2026: Start $0.14 a minute with a $0 platform fee, Build $0.12 a minute plus $299 a month, Enterprise custom. The free tier is the absence of a platform fee, which on Start amounts to much the same thing while you are testing. My verdict is simple arithmetic: stay on Start until you cross 14,950 minutes a month, then move, and re-check the rate on their page before you do because these numbers move.

4. ElevenLabs Agents: minute buckets, and the under-use penalty

ElevenLabs sells minute buckets instead of per-minute usage, which makes it the easiest invoice here to read and the easiest one to waste money on. Six tiers as of October 2026: Free at $0 for 15 minutes of calls and 4 concurrent, Starter at $6 a month for 75 minutes and 6 concurrent, Creator at $22 a month for 275 minutes and 10 concurrent, Pro at $99 a month for 1,238 minutes and 20 concurrent, Scale at $299 a month for 3,738 minutes and 30 concurrent with 3 workspace seats, and Business at $990 a month for 12,375 minutes and 40 concurrent. The workflow builder, knowledge base and multilingual widget start on the free plan, and the commercial licence arrives at Starter.

Those minute counts are odd on purpose. 1,238, 3,738, 12,375. They are what the page says, and they are presumably a conversion from a character or credit allowance underneath, which is a useful reminder that these allowances are conversions, and conversions get revised.

The 1,000 minute workload lands on Pro at $99, with 238 minutes to spare.

Buckets reward an accurate forecast and nothing else

Finish all 1,238 Pro minutes and you paid $0.080 a minute, the second cheapest all-in rate in the table. Use 250 of them and you paid $0.396 a minute, which is worse than every self-serve rate in the table. Nothing about the service changed between those two sentences. The only variable is how well you predicted your own call volume a month in advance, which is a harder forecasting problem than it looks for an inbound line.

ElevenLabs' page showed a first-month discount of 83% on 4 October, taking Starter to $1 and Creator to $11, marked "Until Oct 18". Treat the list price as the real price. A promotion that expires in a fortnight should not decide a platform you will live on for a year.

Best for a team already on ElevenLabs voices with volume that is low and, more importantly, predictable. Pricing as of October 2026 runs from the free plan through Starter $6, Creator $22, Pro $99 and Scale $299 to Business $990 a month. Size the bucket to the volume you can prove, because overflow is a known cost and an unfinished bucket is a silent one.

5. Synthflow: an enterprise entry point, and a page I could not fully read

I am including Synthflow because the SERP includes it and because the one number on its pricing page is a genuinely useful data point. I am also going to be straightforward about the limits of what I saw.

The figure in the page text: enterprise contracts start at $30,000 annually, with final pricing scoped around call volume, concurrency and telephony setup. Divide by twelve and you get about $2,500 a month before anyone has scoped anything, which is roughly 18x the Bland Start bill for the same 1,000 minutes.

Why I am not quoting a self-serve number

Synthflow's self-serve tiers are rendered in JavaScript and did not load when I read the page on 4 October 2026. That means unknown, and unknown is not the same as absent. I am not going to tell you Synthflow has no self-serve plan, because I did not see enough of the page to say that, and I am not going to repeat a figure from a comparison blog that may have read a different version of it. If self-serve Synthflow pricing matters to your decision, ask them and get it in writing.

Best for

An organisation with a procurement process, call-centre volume, and a preference for having the deployment scoped by the vendor.

Pricing

Enterprise contracts start at $30,000 annually per their pricing page, scoped around call volume, concurrency and telephony setup. As of October 2026.

Free tier

Not verified. The self-serve section of the page did not render for me.

Verdict

Budget from $30,000 a year and go in expecting a scoping call. At 1,000 minutes a month this is the wrong shape of vendor, by a factor of about eighteen.

The telephony layer underneath

Every voice agent platform sits on a phone network, and that network has its own price list. Two of the five handle it for you at no extra charge on their own pages: Vapi lists its telephony and SIP as free alongside Vapi WebSockets and Daily WebRTC, and Retell's itemised sample build shows telephony at $0.00 a minute. Bland passes telephony through inside its bundled rate.

If you do bring your own carrier, Twilio is the default. From twilio.com's US voice pricing page, which states its pricing is current as of August 2026, two months before I read it:

• Local numbers: $0.0085 a minute to receive, $0.0140 a minute to make.
• Toll-free: $0.0220 a minute to receive, $0.0140 a minute to make. Receiving on toll-free costs about 2.6x receiving on local, which is the kind of line that quietly reshapes an inbound business case.
• Browser or app, and SIP interface: $0.0040 a minute in both directions.
• Trial: $15 to buy a number and test.

On the component path, Twilio inbound is roughly 20% of Vapi's low-end per-minute cost and about 9% of its high-end cost. Small enough to ignore in a pitch deck, large enough to matter in a margin calculation at scale. If your product makes calls instead of answering them, note that making a call costs about 1.6x receiving one on a local number, which adds about $6 to the thousand on Vapi's low end.

Which should you choose?

Start from what you are optimising, because every one of these platforms is the right answer to a different question.

Lowest possible cost, and you are willing to assemble and maintain it: Vapi. You buy at component cost and pay the rest in your own time. Decide the model first, because that one choice moves the bill more than 2x.

You want one number that is true: Bland. The rate bundles the stack and telephony passes through. Stay on Start until you cross about 14,950 minutes a month, which is where the $299 Build fee finally earns back its $0.02 a minute discount. At 1,000 minutes, Build costs $419 against Start's $140 for identical work.

You need to ship this week with real margin and no contract: Retell. Zero to start, $10 in free credits, 20 concurrent calls, and the component split published so a client can audit your quote.

You already use ElevenLabs voices and your volume is low and predictable: ElevenLabs Agents. Buckets are the simplest invoice in this market. Just do not buy a bucket you will not finish, because 250 minutes on a 1,238 minute plan costs $0.396 a minute.

You have procurement, call-centre scale, and want the thing scoped for you: Synthflow, and budget from $30,000 a year.

You do not want to build any of this: the done-for-you answering services are a different market with a different price floor. Covered separately.

My own bias, stated plainly, because I sell agents and you should discount me accordingly. Price the whole loop, the booking, the CRM record, the follow-up, before you commit to a per-minute rate.

FAQ

Sources