Why Your AI Voice Agent Bill Beats Minutes x Rate

Because minutes × rate is not the formula the vendors bill on. Four mechanics they publish themselves sit in between: Retell’s prompt-token multiplier, which bills a 60-second call as 72 seconds at 4,800 tokens; its 10-second minimum on dynamic openings; Twilio’s per-leg rounding to the next full minute; and Retell’s US$0.10/min concurrency-burst surcharge across the entire call.

None of this is hidden. Every mechanic below is on a vendor’s own documentation or pricing page, in plain language, and was re-read against the raw page on 8 September 2026. The problem is not disclosure. The problem is that a billing-exceptions page is the last document anyone opens before signing, and the first one that explains the invoice afterwards. This page is a neutral walk-through of the arithmetic, not a complaint about any vendor — the mechanics below are, in most cases, defensible cost pass-throughs that the vendor has chosen to publish rather than bury.

The four mechanics that sit between wall-clock minutes and billed minutes

Each one is independent of the others, and each is triggered by a configuration decision rather than by usage. That is the important structural point: your advertised per-minute rate can be constant all month while your effective rate moves by a multiple.

Mechanic	What triggers it	Published effect	Vendor’s own page
Prompt-token duration scaling	An agent whose LLM context exceeds 4,000 tokens	Billed duration = original duration × (prompt tokens ÷ 4,000), rounded up	Retell AI billing exceptions
Minimum billed duration	Calls under 10 seconds that use a dynamic opening message where the AI speaks first	Minimum charge of 10 seconds; Retell’s example bills a 6-second call as 10	Retell AI billing exceptions
Concurrency-burst surcharge	A call that starts while you are above your standard concurrency limit, with burst enabled	US$0.10/min “applied to the entire call duration”, not only the overage portion	Retell AI concurrency & limits
Per-leg carrier rounding	Any voice leg on Twilio Programmable Voice, SIP Interface, Elastic SIP Trunking, Twilio Client or Media Streams	Partial minutes under 60 seconds rounded up to the next full minute, per leg	Twilio minute rounding

A fifth category is not a duration mechanic at all: monthly subscription lines. Retell publishes phone numbers at US$2.00/month, concurrency beyond the free 20 at “$8.00/Concurrency/month”, knowledge bases beyond the first ten at US$8.00 each per month and verified phone numbers at US$10.00 per number per month on its pricing page. Vapi publishes hosting at “$0.05 / min” with model provider costs “At cost ($0 if you bring your own API key)” and concurrency as “10 included + $10 / line / mo”. Those lines do not scale with minutes and therefore never appear in a minutes × rate estimate. If you are still working out which pricing model a vendor is even using, our companion piece on how to read an AI phone agent price page in Australia covers the five quoting structures before the mechanics on this page apply.

How a prompt-token multiplier turns 10 minutes into 30

This is the single largest mechanic on the list, and the least visible, because it is invisible in the call log — the call really did last ten minutes. Retell’s documentation states the rule in two lines:

“Scaling Factor = Prompt LLM Tokens ÷ 4,000”
“Billed Duration = Original Duration × Scaling Factor (rounded up)”

— Retell AI, Exceptions to Our Per-Minute Pricing, Rule 2, read 8 September 2026

The rule is scoped: the page states it applies to “Agents that use more than 4,000 LLM tokens in their prompts”. Below that threshold there is no multiplier. The page’s own worked example is a 60-second call at 4,800 tokens, which gives a scaling factor of 1.2 and a billed duration of 72 seconds.

Citable asset 1: billed-minute multiplier table

This table is a derivation from Retell AI’s published formula, not a measurement of anything we ran. Every row is the vendor’s own arithmetic applied to a constant 10-minute (600-second) wall-clock call, so you can check each line yourself. Substitute your own call length and the multiplier column is unchanged.

Prompt LLM tokens	Scaling factor (tokens ÷ 4,000)	Wall-clock call	Arithmetic	Billed duration
4,000 or fewer	Rule does not apply	10 min 0 s	—	10 min 0 s
4,800 (Retell’s own example, scaled from 60 s)	1.2	1 min 0 s	60 s × 1.2 = 72 s	1 min 12 s
6,000	1.5	10 min 0 s	600 s × 1.5 = 900 s	15 min 0 s
8,000	2.0	10 min 0 s	600 s × 2.0 = 1,200 s	20 min 0 s
12,000	3.0	10 min 0 s	600 s × 3.0 = 1,800 s	30 min 0 s
16,000	4.0	10 min 0 s	600 s × 4.0 = 2,400 s	40 min 0 s
20,000	5.0	10 min 0 s	600 s × 5.0 = 3,000 s	50 min 0 s
32,768 (documented default maximum for the Retell LLM framework)	8.192	10 min 0 s	600 s × 8.192 = 4,915.2 s, rounded up	81 min 56 s

Three things to note honestly. First, the documentation says “rounded up” without stating the unit it rounds to; Retell’s own example produces a whole number of seconds, so we have rounded up to the second and flagged the one row where it matters. Second, the page describes the scaling as applying to billed duration, while the rule is titled “LLM Price Scaling” and justified by model-provider costs. It does not say whether the scaled duration is then applied to every per-minute component or only the LLM line. We show both readings in the reconciliation below rather than pick one. Third, the 32,768-token row is a bound, not a ceiling: Retell’s concurrency page gives 32,768 tokens as the default maximum prompt length accepted when creating or updating a Retell LLM, while the token count the scaling formula reads also includes transcript, tool call history and retrieved knowledge that accumulate during the call. Treat that row as illustrative arithmetic at a documented number, not as a maximum anyone can charge.

The token count is not your system prompt

The most common misreading is that 4,000 tokens is a system-prompt budget you set once. Retell’s page lists what counts toward the calculation: the global prompt, functions (tool descriptions), the state or node prompt, the transcript between agent and user, tool call history and results, and retrieved knowledge base content. Four of those six grow while the call is in progress. A tightly written 2,500-token prompt with three tools can cross the threshold mid-conversation on a long call, and a knowledge-base hit injects retrieved passages straight into the context — which is why mid-call knowledge lookups have a cost profile as well as a latency profile. Retell also names Flex mode as “a common trigger for this rule” because it compiles all node prompts, transitions and tool descriptions into a single context.

The quotable version: a per-minute rate quoted against a 4,000-token agent is not the same product as the same rate quoted against a 12,000-token agent, and only one of them is on the pricing page.

Why every carrier minute rounds up, and why short calls are the expensive ones

Carrier rounding is the oldest mechanic here and the one buyers most often assume has been engineered away. It has not. Twilio’s support documentation states it directly:

“When Twilio’s voice and calling products are billed to your account, any partial minutes (under 60 seconds) are rounded up to the next full minute. For example, an automated one-way call of 1 minute and 20 seconds would be billed as 2 minutes.”

— Twilio, How are Twilio voice and calling product minutes rounded for billing?, last updated 29 May 2026

Two details in that article do the real damage. It names Media Streams in the list of products the policy applies to — and Media Streams is how a voice agent gets audio off the call in the first place. And it states that a two-way call connecting two different Call SIDs for 1 minute and 20 seconds “would bill you for two different calls of 2 minutes”. A separate Twilio article on voice application cost per minute confirms the general rule: “Our prices are charged based on each leg of the call and any partial minutes (under 60 seconds) are rounded up to the next full minute.” Any agent that warm-transfers to a human is a two-leg call, and both legs round independently.

Rounding waste is not proportional to volume. It is worst at the exact call lengths a qualification agent produces.

Wall-clock leg	Billed by carrier	Wasted seconds	Effective rate uplift
0 min 30 s	1 min	30 s	+100%
1 min 1 s	2 min	59 s	+96.7%
1 min 30 s	2 min	30 s	+33.3%
2 min 0 s	2 min	0 s	0%
3 min 10 s	4 min	50 s	+26.3%
5 min 0 s	5 min	0 s	0%

The decision rule that falls out of it: a call that ends one second past a minute boundary costs the same as one that runs the full extra minute, so the cheapest place to end a call is just under the boundary, and the most expensive place is just over it. A voicemail-detection routine that hangs up at 31 seconds and one that hangs up at 61 seconds differ by 100% on the carrier line for identical information.

Australian destinations make this larger, because Twilio’s published Australian rates differ sharply by number type. On Twilio’s Australian voice pricing page, outbound to an Australian mobile is listed at $0.0750/min against $0.0252/min for an Australian local number — roughly three times the rate, on the number type most Australian small-business prospects actually answer. Local numbers are listed at $3.00/month in both of that page’s number tables. The mobile and toll-free monthly figures appear twice on the page and the two tables do not agree with each other — the “Receive calls” table lists mobile numbers at “+ $20.00 / mo” and toll-free at “+ $8.25 / mo”, while the “Phone number pricing” table lists toll-free numbers at $20.00/mo and mobile at $8.25/mo. Read both before you model a number type other than local.

One clarification in fairness to the platforms: Retell’s own pricing FAQ states “Each call is tracked to the nearest second” and “No rounding up per call, no inflated bills.” The per-leg rounding above is a carrier behaviour. It reaches you directly when you bring your own Twilio account — Retell publishes “No Charge for Sip Trunking/Custom Telephony” for exactly that setup — and it is priced into the platform’s own telephony line when you do not.

What concurrency and burst add that has nothing to do with minutes

Concurrency is capacity, not usage, and it is billed as capacity. Retell allocates Pay-As-You-Go workspaces 20 concurrent calls by default and publishes additional capacity at “$8.00/Concurrency/month”. That is a fixed monthly line that is identical whether you place one call or a hundred thousand.

Burst is the mechanic worth reading twice, because the scope of the surcharge is the part people get wrong. Retell’s documentation is unambiguous:

“The burst surcharge applies to the entire duration of any call that started while in burst mode, not just the portion of time spent above the normal limit.”

— Retell AI, Understand concurrency & limits, read 8 September 2026

The same page sets the rate at “$0.10/min surcharge applied to the entire call duration” and defines the burst ceiling as the lower of three times your concurrency limit or your limit plus 300. So a nine-minute call that happened to start during a thirty-second traffic spike carries the surcharge for all nine minutes, including the eight and a half minutes after the spike cleared. The page also says plainly that “Consistent high usage above your normal limit may mean you should raise your base concurrency limit for better cost efficiency” — the vendor itself pointing predictable load at a higher base limit rather than at the burst toggle. If your load pattern is a campaign rather than a spike, the capacity planning in our note on scaling from 100 to 10,000 calls a day is the relevant lever, not the burst toggle.

Two smaller lines from the same vendor’s pricing FAQ round out the picture, and both are honest disclosures rather than traps. On silence: “Billing covers the entire duration of the call because the speech-to-text engine remains active and listening throughout, even during silence.” On voicemail: “For calls that reach voicemail, billing applies only for the duration the AI agent is active on the line” — so a voicemail is not free, it is short.

A worked reconciliation of one example month

Citable asset 2: wall-clock minutes vs billed minutes, itemised

Read the labels before the numbers. Every rate below is a vendor-published figure with a link to the page that publishes it. Every volume, duration and configuration choice is an assumption, marked as such, that exists so you can substitute your own. This is not a measurement, not a typical invoice, not anyone’s real account, and emphatically not Zian pricing. Currency is as each vendor publishes it. Retell and Vapi quote in US dollars. Twilio’s Australian voice pricing page carries no visible currency code, so confirm the billing currency against your own account before you model it — the figures below are used exactly as that page prints them.

Stated assumptions. A1: 1,500 connected outbound calls in one calendar month. A2: mean wall-clock talk time of 3 min 10 s (190 s) per call, giving 285,000 s = 4,750.0 wall-clock minutes. A3: average LLM context of 9,000 tokens across the call, inclusive of transcript and retrieved knowledge. A4: dialling Australian mobiles through your own Twilio account. A5: five Australian local numbers rented. A6: peak concurrency of 60. A7: 15% of calls (225) start while in burst mode during one campaign day.

Published rates used. From Retell’s detailed component pricing: Retell Voice Infra $0.055/minute, Retell Platform Voices text-to-speech $0.015/minute, GPT 5.1 at $0.04/minute — summing to $0.11/minute, which is also the figure Retell’s own on-page estimator returns for that combination. Concurrency beyond the free 20 at $8.00/Concurrency/month. Burst at $0.10/min from the concurrency page. From Twilio’s Australian pricing: outbound to Australian mobile $0.0750/min, local number $3.00/month.

The naive expectation: 4,750.0 minutes × $0.11 = $522.50.

Line	Mechanic in play	Arithmetic	Amount (as published; US$ on the platform lines)
Platform minutes, unscaled (reference only)	None	4,750.0 min × $0.11	$522.50
Platform minutes, as billed	Prompt-token scaling at 9,000 tokens: 9,000 ÷ 4,000 = 2.25	4,750.0 × 2.25 = 10,687.5 billed min × $0.11	$1,175.63
Carrier minutes	Per-leg rounding: 190 s rounds to 4 min, so 1,500 × 4 = 6,000 min against 4,750 wall clock (+26.3%)	6,000 min × $0.0750	$450.00
Number rental	Monthly subscription, no minutes component	5 × $3.00	$15.00
Concurrency capacity	Monthly subscription above the free 20	(60 − 20) × $8.00	$320.00
Burst surcharge	Charged on the entire duration of each bursting call	225 calls × 190 s = 712.5 min × $0.10	$71.25
Total for the example month	—	1,175.63 + 450.00 + 15.00 + 320.00 + 71.25	$2,031.88

Against the naive $522.50, that is 3.89×. The advertised per-minute rate was correct the whole time.

The ambiguity, shown rather than hidden. If the token scaling applies only to the LLM component rather than to the whole per-minute stack — a reading the documentation permits and does not resolve — the platform line is (10,687.5 × $0.040) + (4,750.0 × $0.070) = $427.50 + $332.50 = $760.00, and the month totals $1,616.25 instead. Your invoice detail will tell you which reading applies to your account; the documentation will not. Check it against a single call before you model a year.

Now change only the configuration, not the price. Hold every volume assumption identical, but keep the agent context under 4,000 tokens so Rule 2 never fires, leave burst disabled, and stay inside the free 20 concurrent calls:

Configuration	Platform	Carrier	Subscriptions	Burst	Month total (as published)
Heavy: 9,000-token context, 60 concurrency, burst on	$1,175.63	$450.00	$335.00	$71.25	$2,031.88
Lean: under 4,000 tokens, 20 concurrency, burst off	$522.50	$450.00	$15.00	$0.00	$987.50

Between those two rows the advertised per-minute rate never changes, the call volume never changes, and the invoice roughly doubles — 2.06×, a difference of $1,044.38. That is the real finding: on a per-minute platform, most of your bill is set by architecture decisions, not by the rate you negotiated. Carrier rounding is the one line that barely moves, because you cannot configure your way out of a minute boundary.

What to check before you sign, in the order that matters

All of this is answerable in about twenty minutes, entirely from documents the vendor already publishes. It is worth doing in this order, because the first item is worth more than the rest combined.

1. Find the billing-exceptions page. It may not be reachable from the pricing page: on the Retell pages we read for this article, the pricing page links to the documentation home and the compliance page, and the billing-exceptions page is instead linked from the concurrency documentation. Search the docs for “exception”, “minimum”, “surcharge” and “scaling”. If there is no such page, ask for the equivalent in writing before signing rather than concluding one does not exist.
2. Get your agent’s actual token count, not your prompt’s. Ask for the context size at the end of a long call, with tools, transcript and retrieved knowledge included. That single number sets the multiplier.
3. Ask whether a duration multiplier applies to the full per-minute stack or to the model line only. This is a factor-of-two question and it is not in any documentation we read.
4. Establish who bills the carrier leg. If it is your own carrier account, per-leg rounding lands on you directly; if the platform resells it, ask whether the rounding is passed through.
5. Count the legs, not the calls. Every transfer to a human is a second leg that rounds independently.
6. Price your concurrency at peak, then check the burst policy’s scope. Whether a surcharge covers the whole call or only the overage is a large multiple on a campaign day.
7. Reconcile one real call by hand in week one. Not a month. One call, against the invoice detail, before the pattern is baked into a forecast.

Step one generalises well beyond billing: on voice platforms the marketing page states the price and the documentation states the mechanics, which is the same evidence hierarchy we apply when assessing any vendor claim against its own documentation. If a fact only exists on the marketing page, it is a position, not a specification.

Frequently asked questions

Why is my AI voice agent bill higher than my minutes times the per-minute rate?

Almost always because billed duration is not wall-clock duration. The four published causes are prompt-token scaling of billed duration, minimum billed durations on very short calls, per-leg carrier rounding to the next full minute, and surcharges such as concurrency burst that are charged across a whole call. Monthly subscription lines for numbers and concurrency then sit outside the minutes calculation entirely.

Does the vendor round each call up to the next minute?

That depends on which vendor and which leg. Retell’s pricing FAQ states “Each call is tracked to the nearest second” and “No rounding up per call”. The carrier is a different question: Twilio’s support documentation states that “any partial minutes (under 60 seconds) are rounded up to the next full minute” and applies that policy to Programmable Voice, Twilio Client, SIP Interface, Elastic SIP Trunking and Media Streams. Its worked example bills a 1 minute 20 second call as 2 minutes. If you bring your own carrier account, that rounding is yours.

Does a concurrency burst surcharge apply only to the overage minutes?

Not on the one platform we could find documenting it. Retell’s concurrency documentation states that the surcharge “applies to the entire duration of any call that started while in burst mode, not just the portion of time spent above the normal limit”, at a published rate of $0.10/min. We also read Vapi’s pricing page, which publishes concurrency as “10 included + $10 / line / mo” and states nothing about a burst surcharge, and Retell’s own pricing page, which prices concurrency but does not mention burst. We could not locate an equivalently explicit statement either way on those pages, so ask rather than assume.

What counts toward a prompt-token threshold?

More than the system prompt. Retell’s billing-exceptions page lists the global prompt, functions or tool descriptions, the state or node prompt, the transcript between agent and user, tool call history and results, and retrieved knowledge base content. Because the transcript and tool history grow during the call, an agent that starts under a threshold can cross it mid-conversation.

Am I billed for silence, hold time and voicemails?

Generally yes for silence, and partly for voicemail. Retell’s pricing FAQ states that billing “covers the entire duration of the call because the speech-to-text engine remains active and listening throughout, even during silence”, and that for calls reaching voicemail “billing applies only for the duration the AI agent is active on the line”. Failed connections are stated as not billed.

Are vendors hiding these mechanics?

On the evidence we could find, no. Every mechanic on this page is on a public documentation or pricing page, written in plain language, with worked examples supplied by the vendor. The gap is procedural rather than ethical: pricing pages are read during evaluation and billing-exception pages are read after the first invoice. Reversing that order is the entire fix.

All vendor pages linked above were opened and read in raw form on 8 September 2026. Published prices change; re-check each vendor’s own page before relying on a figure. Currency is as published by each vendor, and Twilio’s Australian voice pricing page carries no currency code — confirm it against your own account.