The cost of a voice minute.
Voice is the one agent workload where the bill is a clock, and the clock runs on four meters at once while the token meter — the only one anybody optimises — is the smallest of the five. That inversion decides every engineering argument downstream: the seconds you spend on a confirmation turn cost single-digit cents and the escalation they prevent costs dollars, so the unit to manage is cost per resolved call, and containment moves it several times harder than the rate card does.
Step 1Four meters on one clock, and a fifth on tokens.
A text agent's invoice is denominated in work: tokens in, tokens out, tool calls, sandbox seconds. A voice agent's invoice is denominated in wall-clock time, and the same wall-clock second is billed simultaneously by almost every layer of the stack.
# What one minute of connected call bills, in parallel carrier / SIP trunk per minute # PSTN leg, inbound and outbound priced apart platform / orchestrator per minute # the hosted voice runtime speech-to-text per minute # streaming ASR, usually billed on audio duration text-to-speech per character # or per second of synthesized audio model per token # the only line that is not a clock
Through October 2026 the published platform fees for the main hosted voice stacks sat roughly in the five-to-fourteen cents per minute band, with the figures at the low end excluding the model and the carrier and the figures at the high end claiming to include them. Treat any single published number as a starting point rather than a quote: the spread between a headline platform rate and a real all-in rate is commonly a factor of two to three, because the headline is one meter and the bill is five. Verify against each vendor's own page before you put a number in a model.
The architectural consequence arrives before the financial one. In text, adding a component adds a cost you can see in that component's line. In voice, adding a serial component lengthens the call and therefore raises four lines at once, including the ones belonging to vendors you did not change.
Step 2The model is the cheap part, which is why the usual optimisation misfires.
Work a single call through with the assumptions written down, because the ratio is the point and the ratio is robust to the exact rates.
# One 4-minute call. Assumptions stated, substitute your own. clock meters 0.12 $/min x 4.0 min = 0.480 model tokens ~20 turns, cached prefix = 0.020 total = 0.500 # Share of the bill by lever: clock 0.480 / 0.500 = 96% tokens 0.020 / 0.500 = 4% # Halving the model's token price saves 0.010 — two percent. # Removing one 30-second turn saves 0.060 — twelve percent.
This is the opposite of the text case, where context growth and reasoning tokens dominate and the standing advice on fewer tokens versus cheaper tokens is about which of those two levers to buy. In voice, both of those levers act on four percent of the bill. A cheaper, faster, smaller model is still worth having — it shortens the turn, which shortens the call, which bills on the clock — but it is worth having for the second reason, not the first, and teams that pick it for the first reason routinely take an accuracy loss for a two-percent saving.
Step 3Silence bills at the same rate as speech.
Per-minute meters do not know whether anybody is talking. The caller thinking, the agent's hold phrase, the dead air while a backend call completes, the greeting, the hold music, the pause an endpointing timer adds to every single turn — all of it is billed at the connected-call rate by every clock meter in the stack.
That makes the fixed silence timeout a line item rather than a UX setting. A 700 ms endpointing threshold on a call with twenty turns is fourteen seconds of billed silence per call, paid whether or not the caller had finished. At twelve cents a minute that is under three cents, which sounds like nothing until you notice it is larger than the entire model bill above it — and that a model-based detector reclaims most of it without a rate negotiation. The mechanism is on end-of-turn detection.
Here is the distinction that resolves the apparent conflict with everything else written about voice latency. Within-turn latency is a perception cost; call duration is a cash cost; they are different levers and they do not always move together. Covering a two-second tool call with “one moment, pulling that up” improves the first and worsens the second. Shortening the conversation by one exchange improves the second and may do nothing for the first. The perception side is priced in pricing latency; the engineering side is the latency budget. This page is only about the cash.
Step 4Quality ceremony has an exact price, and it is small.
Voice is unusual in that you can put a precise figure on a correctness intervention, because the intervention is measured in seconds and seconds have a published rate. The τ-Entity benchmark supplies the cleanest worked case available: a scaffold that prescribes spelling, read-back, correction and confirmation raises robust Pass3 on spoken-entity capture by 14 to 31 points, at a cost of 21 to 28 seconds per call.
# Is the confirmation scaffold worth it? Assumptions stated. scaffold cost 28 s x 0.12 $/min / 60 = 0.056 # paid on EVERY call # What one failed capture costs: second agent call = 0.500 human handling, 6 min at 35 $/h loaded = 3.500 total = 4.000 # paid on failures only # Break-even prevention rate: 0.056 / 4.000 = 1.4%
The scaffold has to prevent a failure on roughly one call in seventy to pay for itself, and the benchmark measures it preventing one in seven at the low end of its range. The margin is not close, and it is not close by two orders of magnitude on the dimension teams actually argue about. The reflex that says “don't add a confirmation turn, it makes the call longer” is pricing the numerator and ignoring the denominator.
Before cutting any turn for speed, compute the two numbers above for your own rates. Almost every correctness turn in a voice flow clears a break-even under three percent, which means the question is never whether the ceremony is affordable — it is whether the ceremony actually prevents the failure, which is an evaluation question and not an economics one.
Step 5The unit is cost per resolved call, and escalation dominates it.
Cost per minute is a rate, not a unit, and a rate cannot be managed because it has no outcome in it. The unit that can is the one agent unit economics argues for everywhere else, instantiated for voice: total cost divided by calls actually resolved, with the escalation path loaded into the numerator.
# Cost per resolved call. Agent call 0.50, escalated call 4.00. containment 60%: 0.50 + 0.40 x 4.00 = 2.10 containment 75%: 0.50 + 0.25 x 4.00 = 1.50 # -29% # Versus halving every clock meter you pay, at 60% containment: 0.26 + 0.40 x 4.00 = 1.86 # -11%
Fifteen points of containment is worth about three times what halving your entire rate card is worth, and it is the cheaper of the two to obtain because it does not require a vendor to agree to anything. The arithmetic also explains a pattern that looks irrational from outside: deployments that add turns, slow down, confirm more and still get cheaper. They moved the denominator.
Two corrections to apply before quoting the number. Escalations are not uniform — a transfer that lands warm, with context, costs materially less human time than one that restarts the conversation, which is the economic case inside escalation and warm transfer. And a call the agent “contained” by giving a wrong answer is not resolved; it is a deferred escalation plus whatever the wrong answer cost, which is the subject of the cost of being wrong.
Step 6What to measure, and the trade you are actually making.
Four instruments, in the order they pay off.
- Billed seconds per call, decomposed. Speech, silence, tool waits, hold. Until silence is a separate number you cannot see the largest reclaimable line in the stack.
- Containment rate and escalation cost, together. Either alone is unactionable; the product of the two is most of your bill.
- Cost per resolved call at p95, not the mean. Call duration is long-tailed for the same reasons agent task cost is, and the tail is where the human time lives. See forecasting agent spend.
- The per-meter rate sheet, dated. Five vendors, five renewal dates, five chances for a unit to change without a deploy. Record which meter each line came from.
The honest trade is that in voice you cannot buy quality with money, only with time — and time is the thing the bill is denominated in, so every correctness decision is a direct, computable purchase rather than an argument about engineering taste. That is a gift, not a constraint: it is the one place in agent engineering where “is this worth it?” has an arithmetic answer, and the answer is usually yes.