What a call costs, and what controls the cost

The platform pays its voice provider for each second that an agent is on a call. Most calls reach an answering machine. The cost risk is in those calls, and not in the conversations with people. The rates on this page were measured on production calls in October 2026. The outcome mix is still an expected range, and not a measurement. A calibration batch of real calls replaces the ranges with measurements for each client.

How the platform is billed

One call is one attempt to reach one person. This page uses the word "call" for one attempt, and the platform pays for the seconds of that attempt:

The voice provider measures each call to the nearest second.
The voice provider does not round up to a whole minute. It does bill at least 10 seconds when the agent speaks first.
A call that reaches an answering machine is billed only for the time that the agent is on the line.
The voice service costs about $0.157 for each minute on average, measured on 538 production calls from 9 to 22 October 2026. The rate varied from $0.115 to $0.194.
The telephone connection is a separate bill from the carrier. The carrier rounds each call up to whole minutes, and an answered call costs about $0.009.
The post-call analysis of each call costs about $0.003.
All together, an answered call costs about $0.088.

The rate is a range and not one number. A part of the rate depends on which grade of AI model the agent uses during the conversation.

The technical bit, translated

Retell — the voice service that holds the live conversation. It tracks each call to the nearest second and adds the seconds up at the end of the billing cycle. The platform needs it because the platform does not process real-time audio itself. On production its price has four parts. The voice engine is about $0.055 for each minute (35.5% of the cost). The AI model, gpt-4.1, is about $0.045 (35.3%). A surcharge for a long prompt, over 4,000 tokens, is about $0.06 (19.2%). The speech generation is about $0.015 (9.7%). The connection from Retell to the carrier costs nothing at Retell, because the platform brings its own line.
Telnyx — the carrier that supplies each client's phone numbers. It bills the telephone connection separately: $0.005 for each minute on the public network plus $0.002 for each minute of call control, in whole minutes. That is about $0.009 for each answered call. A local number rents for about $1 a month, and a text message costs about $0.011. The platform needs it because a call must travel over a real telephone network.
OpenAI — reads each finished call and extracts its outcome. That costs about $0.003 for each call. The same provider also does the conversation review of the automatic audit. The review uses a small model, gpt-4.1-mini. One review reads about 1,050 tokens and writes about 140. A token is a short part of a word. These token counts give the cost of 0.06 cents for each review.
The blended rate. This model uses $0.12 to $0.16 for each minute. The measured production rate of $0.157 sits inside that range.

What happens to 1,000 calls

The cost of a campaign follows the mix of outcomes, and not the price of a conversation. For every 1,000 calls, the model expects this mix:

10 to 50 calls reach a person. Each one lasts about 90 seconds.
500 to 800 calls reach an answering machine. The agent ends each one after about 9 seconds. This page calls such a call a machine call.
0 to 200 calls get no answer. Nobody picks up, or the network ends the call. These calls use no agent time.
Across all 1,000 calls, the average billed time is about 12 seconds.

Twelve seconds at $0.12 to $0.16 for each minute gives the cost:

About 2.4 to 3.2 cents for each call.
About $24 to $32 for 1,000 calls.
About $700 to $1,000 for a pilot term of 30,000 calls.

The upper part of the diagram splits 1,000 calls into three groups and shows each group's share of the calls next to its share of the cost. The machine calls are the largest group in both, at 500 to 800 calls. The calls answered by a person are 10 to 50 calls, and the unanswered calls are 0 to 200 calls with no agent time and no cost. The lower part is a table of billed seconds for each outcome, drawn as bars. The bar for the time the agent takes to end a machine call is the longest bar in that table. — Two things to notice. First, the machine calls are most of the calls **and** most of the cost, although each one is short. Second, in the lower table, the longest bar is the time that the agent takes to end a machine call. That bar is the one number on the page that the platform controls directly.

What moves the cost

Four changes move the real cost away from the model. Each item gives the change in the average billed time, and then the change in the cost:

A conversation with a person grows from 90 to 180 seconds, at 3 answers in every 100 calls: +2.7 seconds, +22%.
People answer 5 calls in 100 instead of 3, at 90 seconds each: +1.8 seconds, +15%.
The agent takes 20 seconds instead of 9 to end a machine call, at 65 machine calls in every 100: +7.2 seconds, +60%.
The detection misses 10 machines in every 100, so each missed machine runs to the 45-second silence limit: +2.9 seconds, +24%.

The two large numbers are the two machine lines. Machine calls are 50 to 80 of every 100 calls, so a small change to a machine call multiplies across the whole day of calling.

The plain conclusion:

How fast the platform ends a machine call matters more than how long people talk.
The platform controls both machine lines through the settings of the agent.
The two human lines are properties of the lead list, and the platform does not control them.

What one outcome costs

An outcome is a result that a broker acts on. The calls that reach a person do not all produce one:

5 to 10 of every 100 calls answered by a person produce an outcome.
One outcome therefore needs 200 to 2,000 calls.
The call cost of one outcome is $5 to $64.
That figure covers the calls only. It does not cover the transfer, or the people who handle it.

The range is wide because the lists are different. Which end of the range applies to a client is a property of that client's list, and not a property of the platform. The client dashboard reports the cost for each outcome from the calls that really happened.

The same dashboard also answers the other side of the question: what the calling has been worth so far. Two estimates sit above the outcome counts — an estimated pipeline value and an estimate of rep time saved — both priced from the calls that really happened, against a published set of pricing rules everyone works from. Neither one is the value of deals a client's own underwriting actually funds; that figure happens weeks later, inside a client's own systems this platform never sees, and the platform does not claim to measure it.

A separate Pipeline screen tracks one slice of this forward work on its own: it counts every merchant who agreed to a callback that has not happened yet, by month. It is a count, not a value — the estimated pipeline value above is the only place a dollar figure for this forward work appears.

The controls

Each control below carries one status marker:

(built) — the platform does this today, and it does it automatically.
(proposed) — a plan only. Nothing of this runs today.
(a person does this today) — the platform does not do this. An operator does it by hand.
(open gap) — the platform names the control, but nothing enforces it.

These are the controls, in the order of what already runs:

The platform records what it spends, in a spend ledger. The ledger holds the voice provider's price for every priced call, the carrier's charges for calls, call control and text messages, and the cost of each post-call analysis. The carrier's charges are read every hour, so they arrive a few hours late. The ledger keeps its entries even when a lead is reset. (built) The ledger does not record number rent, number reputation lookups, hosting, the do-not-call list subscription ($449 a month) or the email service. (open gap)
The agent ends a call as soon as it detects an answering machine. This one setting keeps the machine call at about 9 seconds. (built)
A call cannot last longer than 7 minutes. The model expects a conversation of about 90 seconds, so the team considers a shorter limit of 5 minutes. (built)
The silence limit is 45 seconds. The agent ends a call after 45 seconds of silence. (built) The limit is too long: a machine that the detection misses then costs 45 seconds. The proposal cuts the limit to about 15 seconds. (proposed)
A daily spend ceiling exists as a setting for each client, and the platform now enforces it. The ceiling is $400 a day for each client unless it is set otherwise. The platform totals everything in the spend ledger for that client — voice service, carrier and post-call analysis — for each calling day. A calling day runs from 10:00 UTC to 10:00 UTC. The moment the total reaches the ceiling, a spend breaker trips and blocks every further dial for that client until a person resets it. (built) This is a full stop, not a pause on only the newest leads: calls already under way still finish, but nothing new goes out until the reset. The total does not include the costs that the ledger leaves out (see above), and the carrier's charges arrive a few hours late. A stop in the middle of a term leaves leads with part of their calling schedule not made. The day-zero plan of the next version aims to make the spend a limit at the moment a lead starts, rather than a halt partway through the day.
A daily gauge of call length by outcome, for each client. It raises an alert when the measured length is 25% above the model. (proposed)
A calibration batch of 500 to 1,000 real calls for each client, before volume. It costs about $15 to $30, and it replaces the ranges on this page with that client's own measurements. The platform does not do this by itself. (a person does this today) The same batch also measures the client's answer rate, and the number rotation rule needs that rate as its baseline.

Updated 26 October 2026. The rates and the ledger on this page reflect the measurements of 9 to 22 October 2026 and the changes shipped on 25 and 26 October 2026.

What the automatic audit costs

Every 15 minutes, the platform audits every running campaign of a real client. Most of the audit costs nothing:

The checks of the platform's record and of its running jobs cost nothing. They read data that the platform already holds.
The reply-wait measure and the early "are you still there?" measure cost nothing. They use the word timings that the voice service sends with each call.
The conversation review costs money. It is one small AI request for each answered call, and it occurs one time for each call.
One review costs about 0.06 cents for each answered call. That is about 65 cents for each 1,000 leads.

The spend ledger records the cost of each review. Thus the review counts in the client's daily spend total, the same as all other entries.

Each audit has a limit:

One audit reviews a maximum of 25 answered calls for each campaign.
The calls after the first 25 wait for the next audit, 15 minutes later.

One check is not automatic. An engineer makes a second, more accurate transcription of a sample of call recordings, to check what the voice service heard. The engineer does this by hand when it is necessary. The cost of this check is not in the figures above.

One honest note on the data

The only real data so far comes from one client with very high volume. That client has low answer rates, and the people who answer end the calls quickly.

Two consequences follow:

Short calls give a low average, so the model makes the cost look lower than it is.
A client whose calls become real conversations runs longer human calls, and pays more.

The platform therefore calibrates for each client, and not one time for the whole platform. Until a client has its own calibration batch, every figure on this page stays an expected range.