The advertised per-minute price: why a $0.06 headline becomes $0.10 to $0.33 all-in
The headline per-minute price is one layer of seven. How the other layers add up, the billing edge cases that inflate the bill, and how to get an all-in number.
By Voice Agent Bible Research · 5 min read
Last verified 30 Sept 2026v1.0Published 30 Sept 2026
The advertised per-minute price is usually the orchestration fee alone; speech, language model, voice, telephony, numbers and tooling roughly double to quintuple it.
Symptoms a buyer notices
- The pricing page shows one number per minute and a footnote about pass-through costs.
- The first real invoice is several times the headline, with line items you did not budget for.
- The vendor cannot tell you what a transferred, abandoned or silent minute costs.
- Per-minute quotes from different vendors are not comparable because each bundles a different set of layers.
How it shows up
The pricing page shows a single number: five, six or seven cents per minute. Underneath, in smaller type, a footnote says that language-model usage, premium voices and telephony are billed at provider rates. You build a budget on the headline. The first invoice arrives at two to five times that figure, with line items for speech-to-text, text-to-speech, model tokens, inbound minutes, outbound minutes, number rental and, sometimes, a monthly minimum or a concurrency fee.
The second symptom is that quotes do not compare. One vendor bundles the voice and the model into a tiered rate. Another passes everything through at cost. A third sells a subscription with included minutes and an overage rate. A fourth prices per outcome. All four describe themselves as "from" some number of cents per minute, and none of those numbers mean the same thing.
Why buyers care
Voice agents are bought on a cost-per-call comparison with a human, and that comparison is decided by the all-in number, not the headline. A budget that is wrong by a factor of three at the first invoice damages the project's credibility inside the buyer's organisation, and that damage lands on the sponsor, not the vendor.
The excluded layers are also the ones that scale with your call shape rather than with a flat minute. Long calls with many tool calls cost more in model tokens. Calls that end in a transfer keep billing on some layers after the agent stops talking. Outbound campaigns pay for every ring-out and voicemail. None of this is visible in a per-minute headline, and all of it is visible in the third month of production.
The mechanism
A single automated call runs through several billed layers at once. Using the list-price ranges on this site's TCO calculator, reviewed on 2026-09-30, the layers and their observed ranges per connected minute are:
| Layer | Low | Typical | High |
|---|---|---|---|
| Orchestration platform fee (the usual headline) | $0.02 | $0.05 | $0.15 |
| Speech-to-text | $0.004 | $0.008 | $0.015 |
| Text-to-speech (per minute of agent speech) | $0.01 | $0.03 | $0.06 |
| Language model (tokens converted at typical turn lengths) | $0.005 | $0.02 | $0.06 |
| Telephony, inbound leg | $0.0085 | $0.012 | $0.03 |
| Telephony, outbound leg | $0.013 | $0.02 | $0.06 |
| Testing, monitoring and QA tooling (amortised) | $0 | $0.005 | $0.02 |
Phone number rental sits outside the per-minute maths at roughly $1 to $10 per number per month.
Summing the inbound layers gives about $0.125 per connected minute at the typical column and about $0.33 at the high column. That is the arithmetic behind the title: a headline in the $0.05 to $0.07 band becomes roughly $0.10 to $0.33 once the layers that the headline excludes are added at typical to high list prices. Outbound calls add the higher termination leg and land above that.
Two independent public write-ups reach the same shape. One reported component ranges of $0.005 to $0.02 for speech-to-text, $0.02 to $0.10 for the language model, $0.05 to $0.18 for text-to-speech and $0.005 to $0.03 for domestic inbound telephony, giving a fully loaded floor near $0.08 and a ceiling near $0.40 for a premium-voice, model-heavy call, and placed the all-in figures for developer platforms in bands of roughly $0.09 to $0.35 against published platform fees of $0.05 to $0.10. Another framed the "loaded minute" at realistic volume and stated that first invoices commonly run three to five times the homepage price. A third source, from a vendor selling subscription plans, gave the whole market a range of $0.01 to $1 per minute and noted overage rates of 1.5 to 2 times the base rate once included minutes are exhausted.
The layers that move most with your configuration are the model and the voice. Long system prompts, many tool calls per turn and verbose tool results push token cost up quickly. Cloned or expressive voices sit at the top of the text-to-speech range, and one source reported custom-voice premiums of $0.02 to $0.07 per minute or one-time design fees of $1,000 to $5,000. Multilingual support was reported as an add-on starting around $0.01 per minute. Compliance storage, recording retention and premium support are billed outside the minute altogether.
Evidence
The ranges above are list prices observed on public pricing pages and in public cost breakdowns as of the review date; they are estimates for planning, not quotes, and they move every quarter. Where the two reported sources disagree (text-to-speech at $0.01 to $0.06 versus $0.05 to $0.18, for example) both are shown, and the difference is mostly whether premium voices are included in the band.
The billing edge cases below are not in any headline and are the reason the all-in number is a range rather than a point.
- Transfers. When the agent hands a call to a human, the telephony legs continue until the human hangs up, and on some platforms the orchestration fee continues as well because the platform stays bridged. A warm transfer adds a second outbound leg. On a call shape where a quarter of calls transfer, the transferred tail can be a large share of billed minutes that the agent never handled.
- Abandoned and unanswered calls. Outbound attempts that ring out, hit voicemail or are abandoned in the first seconds still incur telephony, often rounded up to a minimum increment. Some components bill a minimum per connection even when no audio flows. Voicemail detection is itself a billed feature on some platforms.
- Silence and hold. Speech-to-text is billed on audio duration, including silence, hold tones and the seconds before the caller speaks. Text-to-speech is billed on what is generated, including sentences the caller interrupted and never heard. The platform fee bills on connected minutes regardless of who is talking.
- Rounding and minimums. Per-minute billing that rounds up makes a 61-second call cost two minutes. Monthly minimums and committed-use tiers change the effective rate at low volumes; a plan with a fixed fee and 200 included minutes works out near $0.50 per minute if you use them all and higher if you do not.
- Concurrency. Some platforms charge for reserved concurrent lines or throttle above a limit; peak-hour traffic can require paying for capacity you use for two hours a day.
- Model and voice changes. Switching to a larger model or a premium voice changes the pass-through rate immediately, and the change appears on the next invoice unless the vendor publishes a pass-through schedule.
How to test for it in a demo
Price is a demo item. Bring your call shape (calls per month, average minutes, transfer rate, outbound share, number of lines) and run three tests.
- Ask for the itemised all-in number. Every layer, typical and high, at your volume, plus a monthly total. Then enter the same call shape into the TCO calculator at /tools/tco-calculator/ and compare. A vendor number below the calculator's low sum means a layer is missing; ask which one.
- Price the edge cases. Ask what a two-minute call ending in a warm transfer costs on every layer until the human hangs up, and what a five-second abandoned outbound attempt costs. Vendors who have run production traffic answer immediately.
- Ask for a sample invoice. A redacted invoice from an existing customer with a similar call shape shows every line item that exists. The line items you did not expect are the finding.
Record the answers next to the headline price. The ratio between the two is a property of the vendor's bundling, and it is the number to carry into the comparison.
Questions to ask vendors
- What is the all-in cost per connected minute for our call shape, itemised by layer, including telephony and the language model? A good answer is a line-item breakdown with typical and high cases and an explicit list of exclusions.
- How are transferred, abandoned and silent minutes billed on each layer? A good answer is per layer, and says whether agent billing stops at transfer and what minimum applies to short calls.
- What happens to the price when we change the language model or the voice? A good answer is a published pass-through schedule.
- What is included in the platform fee, and what is passed through at provider rates? A good answer names each layer explicitly rather than saying "everything you need".
None of this makes the headline price dishonest. It makes it one layer of seven, and the buyer's job is to price the other six before signing.
Questions to ask vendors
- 01
What is the all-in cost per connected minute for our call shape, itemised by layer, including telephony and the language model?
A good answer: A line-item breakdown with typical and high cases, a monthly total at your volume, and an explicit list of what is not included.
- 02
How are transferred, abandoned and silent minutes billed on each layer?
A good answer: A clear statement per layer, including whether agent minutes stop at transfer, whether speech recognition bills during hold, and what minimum applies to short calls.
- 03
What happens to the price when we change the language model or the voice?
A good answer: A published pass-through schedule, so you can see the increment before you switch rather than on the next invoice.
Related
- Best practice
- Best practice
- Use case