AI Voice Agent Cost Per Minute: What a Call Actually Bills, Layer by Layer

Every article about AI voice agents quotes a platform fee. Vapi is five cents a minute. Retell is five and a half. Bland is eleven to fourteen. Those numbers are all correct and none of them is what you will pay.

A voice agent call bills across five separate layers, and only one of them is the platform. The other four are speech-to-text, the language model, text-to-speech and the phone line. Some platforms bundle two or three of those into the headline rate. None of them bundles all five. So the useful question is not "which platform is cheapest" but "what does a minute of assembled call actually cost, and which layer is worth arguing about".

This post does the arithmetic. Every rate below was fetched from the vendor's own pricing page on 13 August 2026, and every assumption behind the maths is stated so you can substitute your own numbers. If you want the head-to-head platform comparisons instead, they are in Vapi vs Retell AI and Best AI Voice Agent Platforms in 2026. This one is the cost model those two sit on top of.

The finding, before the working

Telephony costs more than the AI.

On a sensibly chosen stack, speech-to-text, the language model and the voice synthesis together come to about 1.4 cents a minute. The phone line comes to about 1.8 cents a minute. Most of the effort that goes into costing a voice agent goes into comparing model prices, and almost none of it goes into the carrier bill, which is the larger of the two.

The platform fee is larger than either, at five to twelve cents. But the platform fee is the one number you cannot negotiate or optimise, only choose once. The layers you can actually move are smaller than people assume, with one exception covered further down.

The assumptions

Cost per minute is meaningless without a call shape, so here is the one used throughout. Change any of it and the numbers move.

• A conversational call where the agent speaks roughly half the time. A scripted outbound pitch skews higher, a qualification call where the caller does the talking skews lower.
• Speech at about 150 words a minute, roughly six characters a word including the space. That is about 900 characters a speaking minute, so 450 characters of synthesised speech per call minute.
• About four conversational turns a minute.
• Around 2,000 tokens of context per turn, being the system prompt plus the transcript so far, and about 60 tokens out. That works out at 8,000 input tokens and 240 output tokens per call minute.
• Outbound calls to United States local numbers, on Twilio's published list rates.

The transcript grows as the call proceeds, so a ten-minute call carries more context per turn than a two-minute one. Treat the token figures as an average across a three-minute call rather than a constant.

Layer one: speech-to-text

Deepgram's pricing page is the clearest published card in this category, and most platforms route to it by default.

Read the small print on those. The page marks the streaming rates as limited-time promotional. Nova-3 Monolingual reverts to $0.0077, and Flux English to $0.0077 as well. If you are costing a contract that runs into 2027, use the standard rates, not the ones on screen today. At 10,000 minutes a month that difference is about $29, which is small, but it is the pattern that matters: several of the figures on these pages have expiry dates attached.

OpenAI is competitive here if you are already in that account. The gpt-4o-mini-transcribe model is billed at an estimated $0.003 a minute and gpt-transcribe at $0.0045, per OpenAI's pricing page. The live-transcription models are much dearer at $0.017 a minute, so check which endpoint your platform actually calls.

Add-ons bill separately and quietly. PII redaction is $0.0020 a minute, keyterm prompting $0.0013, entity detection $0.0017. A healthcare or financial agent that needs redaction on every call has just added forty per cent to its transcription line.

Use $0.0048 a minute.

Layer two: the language model

This is the only layer that can swing by two orders of magnitude, and it is the one most people set and forget.

Holding the call shape constant at 8,000 input and 240 output tokens a minute, against OpenAI's current rates:

The cheapest and dearest rows differ by a factor of about ninety. Put another way, running a flagship model on a voice agent costs roughly what the entire Vapi platform fee costs, on top of the platform fee.

Two things follow. First, the model is the only layer where a bad default is genuinely expensive, so it is the one worth an hour of testing. Second, most of what a voice agent asks a model to do is short, structured and heavily constrained by the system prompt, which is exactly the workload the small models handle well. Very few receptionist or qualification agents need a flagship.

Prompt caching is the free win here. Voice agents resend the same system prompt on every single turn, which is the textbook case for it. If 6,000 of those 8,000 input tokens are a cached system prompt, the gpt-5-mini cached input rate of $0.025 per million takes the minute from $0.0025 to about $0.0011. Roughly half, for a configuration change. OpenAI's caching guide covers the 1,024-token minimum and where the breakpoints fall. If you want to model your own prompt shape rather than mine, the OpenAI API cost calculator does the token arithmetic.

Use $0.0025 a minute for a small model without caching, which is the conservative figure.

A note on speech-to-speech models

Realtime models such as gpt-realtime-mini collapse transcription, reasoning and synthesis into one bill, at $10 per million audio input tokens and $20 per million audio output tokens against $32 and $64 for the full gpt-realtime. Whether that lands cheaper than an assembled pipeline depends entirely on how many audio tokens a minute of speech produces, which is a function of the audio tokeniser and not something worth guessing at. If you are considering that route, run a hundred real minutes and read the bill. It is the only honest way to get that number, and anyone quoting you a confident per-minute figure for realtime audio has probably not measured it either.

Layer three: text-to-speech

Voice is billed per character almost everywhere, which is why it has to be converted before it can be compared.

Retell publishes its voice layer already converted, at $0.015 a minute for platform, Minimax, Fish, Cartesia and OpenAI voices, and $0.040 a minute for ElevenLabs voices, per Retell's pricing page. Note what that implies: choosing an ElevenLabs voice costs about two and a half cents a minute more than the alternatives, which is more than the transcription and the model put together. That is a real trade, and it may well be worth it if the voice is the product. It is worth making deliberately rather than by picking the nicest demo.

The Flux TTS row is the one to watch. It is free right now and $0.0450 per thousand characters from 13 September 2026, which would take it from nothing to two cents a minute overnight. Anyone building a cost model on today's page needs a diary entry.

Use $0.0068 a minute.

Layer four: the phone line

This is where the money quietly is, and where nearly every published comparison stops at the first row of the table.

From Twilio's United States voice pricing:

Media Streams is the line nobody mentions. An AI voice agent has to get the audio out of the call and into the platform, and on Twilio that is a bidirectional media stream billed at $0.0044 a minute on top of the call itself. So the real telephony cost of an outbound AI call is not $0.0140 but $0.0184. That single overlooked row is close to a quarter of the layer, and it is larger than the entire transcription bill on its own.

Three practical consequences.

Toll-free is the expensive default. An inbound toll-free minute is $0.0220 against $0.0085 for a local number. Businesses reach for toll-free out of habit and more than double their telephony line to do it. If the caller is local and dialling from a mobile with inclusive minutes, which is most callers, the toll-free number is buying you very little.

Inbound is cheaper than outbound, by $0.0055 a minute. An inbound receptionist agent and an outbound calling campaign are not the same cost model, which is why outbound AI calling needs its own budget rather than a copy of the receptionist one.

Recording storage compounds. At $0.0005 per minute per month, a minute recorded in January is still billing in December. Record everything at 10,000 minutes a month and by month twelve you are storing 120,000 minutes, or about $60 a month, rising. It is not a large number but it is a permanently rising one, and a retention policy fixes it in an afternoon.

If you want to model a specific mix of numbers, directions and add-ons, the Twilio cost calculator handles it, and Twilio voice setup for AI agents covers wiring the number to the platform.

Use $0.0184 a minute.

Check how your carrier rounds

Per-minute billing rounded up to the whole minute turns a twenty-second call into a sixty-second charge. On an inbound receptionist that barely registers. On an outbound campaign with a high hang-up rate it is the single largest gap between the model you built and the invoice you receive, because the calls that fail fastest are billed at three times their length. Pull one real month of call detail records and compare billed minutes against connected seconds before you trust any forecast, including this one.

Layer five: the platform, and the total

Now the assembled figure. Same call shape, outbound, US local, own provider keys where the platform allows them.

A few things this table earns.

Deepgram's own agent API undercuts the orchestration-only platforms, because at $0.050 a minute it includes the transcription and the voice that Vapi charges you separately for. That rate is promotional too, reverting to $0.065 on 13 September 2026, which would put it at $0.0859 and just behind Vapi. Worth knowing before you commit, and worth revisiting in a month.

The spread across the sensible options is about two cents a minute. Deepgram at $0.0709 against ElevenLabs at $0.1009 is a real difference, but it is not the order-of-magnitude difference the marketing implies. At 3,000 minutes a month it is ninety dollars.

Bland's bundle is the outlier, but bundling is what it sells. One rate, no provider accounts, no key management. If nobody on the team wants to own five vendor relationships, that premium is a salary saving rather than a cost. Bland's pricing page sets out the tiers, and its daily call caps matter more than its rate, which the platform roundup goes into.

What a call costs

At $0.0825 a minute, a three-minute outbound call costs about 25 cents. An inbound one on a local number costs about 23 cents, because inbound minutes are cheaper.

Scaled out on the Vapi stack:

Then add the things that do not bill per minute.

Concurrency. Ten thousand minutes a month at three minutes a call is about 3,300 calls, roughly 111 a day. Averaged across a working day that is barely one concurrent line, but averages are not what break. If your peak needs twenty simultaneous calls, Retell includes twenty and Vapi includes ten then charges $10 a line a month, so the Vapi bill of $825 becomes $925 while Retell stays at $880. The cheaper per-minute rate loses on peak capacity. This is the same pattern the Vapi and Retell comparison found from the other direction.

Numbers. $1.15 a month each for local, $2.15 for toll-free. Trivial until you are running area-code-matched outbound and holding two hundred of them, at which point it is $230 a month and nobody budgeted for it.

Answering machine detection. $0.0075 a call. On outbound where a large share of dials hit voicemail, that is $25 on 3,300 calls, and it is money well spent because the alternative is paying for the agent to talk to an answerphone.

Which layer to actually optimise

In order of what returns the effort.

One: pick a small model and turn on prompt caching. This is the only change that can move the bill by tens of per cent, it takes an afternoon, and for most agents it costs nothing in quality.

Two: fix the phone side. Move off toll-free where you can, put a retention policy on recordings, and check the rounding. Bringing your own carrier drops the Twilio share to $0.0040 a minute for trunking plus $0.0044 for the media stream, and puts the actual call minutes onto wholesale rates. That is worth doing somewhere above ten thousand minutes a month, and not worth the setup below it. Retell charges nothing for telephony if you connect your own SIP trunk, which makes it the cheapest of the group for anyone who already has carrier relationships.

Three: choose the voice deliberately. ElevenLabs at $0.040 a minute against $0.015 for the alternatives is two and a half cents, which is more than the model and the transcription combined. Sometimes that is exactly the right purchase. Just make it knowingly.

Four, and last, the platform. It is the biggest single line and the one you can do least about after the fact, which is why it deserves care at the decision and none of your attention afterwards. Two cents a minute separates the reasonable options. Migrating between them to chase that will cost more in engineering time than it saves in a year.

Where these numbers came from

Every figure here was read off a vendor pricing page on 13 August 2026: Deepgram, Twilio, OpenAI, Vapi and its cost routing documentation, Retell, ElevenLabs Agents and Bland.

Several of them carry expiry dates. The Deepgram streaming transcription rates and its Voice Agent tiers are both marked promotional, and Flux TTS becomes chargeable on 13 September 2026. Check the source pages rather than this article before you commit budget, and never lift a competitor's rate from a vendor's own comparison table. The Bland FAQ, for instance, describes Retell's voice infrastructure as about $0.07 a minute when the Retell card says $0.055. Everyone's comparison page flatters its author.

To run these layers against your own call shape rather than mine, use the voice AI cost per minute calculator, or the AI receptionist cost calculator if you are comparing against the cost of a person answering the phone. If you are building the thing rather than costing it, how to build an AI receptionist walks through the setup, and the Retell AI review covers that platform's layered card in detail.

Frequently asked questions

What does an AI voice agent cost per minute?

Between about 7 and 14 cents a minute all in, on the assumptions in this post. A sensible assembled stack on Vapi comes to $0.0825 a minute, the Deepgram agent API to $0.0709, Retell to $0.0880 and the Bland Build tier to $0.1384 plus a $299 monthly fee. A three-minute call therefore costs roughly 21 to 42 cents depending on the platform. Your figure will differ mainly on how much the agent talks and which model you run.

Why is telephony more expensive than the AI?

Because the AI layers have been in a price war for two years and the phone network has not. Transcription, a small model and a decent voice together come to about 1.4 cents a minute. An outbound US call is $0.0140 plus $0.0044 for the media stream that carries the audio to your platform, or $0.0184. The media stream is the part almost every comparison omits, and it is close to a quarter of the telephony line on its own.

Which layer should I optimise first?

The model, then the phone line. The model is the only layer that can swing by two orders of magnitude, from $0.0005 a minute on gpt-5-nano to $0.0472 on a flagship, and prompt caching roughly halves whatever you land on because voice agents resend the same system prompt every turn. After that, move off toll-free numbers, set a recording retention policy, and look at bringing your own carrier once you are past about ten thousand minutes a month.

Does a cheaper model make the call worse?

Not usually, for this workload. Most voice agents do short, structured, heavily constrained tasks: qualify a caller, book a slot, answer from a fixed knowledge base. That is what the small models are good at. The place a larger model earns its cost is open-ended conversation with genuine reasoning, which is rare in a receptionist agent and common in a sales one. Test both on your own transcripts rather than assuming either way.

Is it cheaper to bring my own Twilio account?

Almost always, and more so at volume. Platforms that resell telephony add their own rate on top, whereas your own account gets you list price and the option of wholesale rates through BYOC trunking at $0.0040 a minute. Retell charges nothing at all for telephony if you connect your own SIP trunk. The catch is that the phone numbers, the compliance and the carrier relationship become yours to manage, which is real work. Below a few thousand minutes a month it is rarely worth it.

How accurate are these figures?

The rates are exact and dated. The per-minute conversions are arithmetic on stated assumptions, which is the part to check against your own calls: how much the agent speaks, how long the transcript grows, how many turns a minute. The single largest source of error in practice is not the rates at all but call-length rounding on short calls, so pull a month of real call detail records before you trust any forecast.

Costing a voice agent properly before you build it is most of the work of building it profitably. If you would rather have someone assemble and price the stack for you, I take voice agent builds through Fiverr and Upwork, and I publish build walkthroughs on YouTube.