What a lead costs
A call costs a few cents. A lead costs much more, because the platform calls each lead up to 15 times. This page gives the rates measured on production from 9 to 22 October 2026. It shows the arithmetic from 1 call to 1 lead, and the totals for 10,000 to 100,000 leads. Each input is either measured from real calls or modelled as an estimate.
Why a lead costs much more than a call
One call is 1 attempt to reach 1 person. A lead gets many attempts before the platform stops:
- The calling cadence allows at most 15 attempts per lead: 2 a day for 7 days, then 1. The cadence runs once, with no rest.
- An answering machine does not end the lead. The lead's journey continues to the next attempt.
- The average lead gets approximately 12 attempts. This is modelled: at 3 answers per 100 calls, the platform never reaches 63% of leads in 15 tries.
Thus the cost of a lead is the cost of 12 to 15 calls. Most of those calls reach a machine.
The real rates
LeadRefinery measured the rates on 538 production calls from 9 to 22 October 2026, where a measurement was possible. The rest is modelled. Retell is the voice service that holds the conversation. Telnyx is the telephone carrier. OpenAI reads each finished call. The table shows time in seconds (s) and says which inputs are measured:
Updated 26 October 2026. An earlier version of this page priced a 16 s machine call at $0.053. The correct figure at about $0.16 per minute is $0.043 (16 ÷ 60 × $0.16). At the measured $0.157 per minute it is $0.042. Every total below is recalculated from the corrected figure and the October measurements.
| Input | Value | Measured or modelled |
|---|---|---|
| Retell, production agent | $0.157 per minute | Measured: 538 priced production calls, 9 to 22 October 2026. The rate varied from $0.115 to $0.194 per minute. The production agent pays about $0.055 for the voice engine, $0.045 for the AI model, $0.06 for a long-prompt surcharge (the prompt is over 4,000 tokens) and $0.015 for speech, per minute. Retell bills per second while the agent is on the line, with a minimum of 10 s when the agent speaks first. |
| Retell, staging agent | $0.30 per minute | Measured. The staging agent adds personal-data scrubbing, colloquial mode and a guardrail. Do not copy that configuration to production before you read this page. |
| A call that reaches a person | 29 s, $0.076 | Measured: the average answered production call, 9 to 22 October 2026. These include QA persona calls. A real conversation can take 90 s ($0.24). |
| A call that reaches a machine | 16 s, $0.042 | The 16 s was measured on 9 production calls only. 16 s at $0.157 per minute is $0.042. What a call costs uses 9 s ($0.024) as the target. If the agent speaks first, Retell bills at least 10 s ($0.026). |
| Telnyx minutes | $0.007 per minute, about $0.009 per answered call | $0.005 per minute for the public telephone network plus $0.002 per minute for call control, billed in whole minutes. The leg from Telnyx to Retell costs nothing. An unanswered call costs nothing. |
| OpenAI post-call analysis | about $0.003 per call | The model reads each finished call and extracts its outcome. This page applies it to every answered call. |
| Telnyx number rent | $1.00 per number per month | About $1 for a US local number; the carrier's price list says $0.79. |
| Attempts per lead, maximum | 15 | The cadence: 2 a day for 7 days, then 1. The cadence runs once, with no rest. A machine answer does not end the lead. |
| Attempts per lead, average | approximately 12 | Modelled: at 3 answers per 100 calls, the platform never reaches 63% of leads in 15 tries. Production has no real cadence campaign yet. Its 7,220 attempts come from the QA persona campaign, 131 attempts per persona. Thus this input is a model, not a measurement. |
| Outcome mix per 100 calls | 65 machine, 3 person, 32 no answer | Modelled, from What a call costs. Not yet measured on real leads. |
| Dials per number per day | 75 | The daily dial cap of the platform, per phone number. |
| Reserve numbers | 25% of working | Spare capacity for number rotation. |
The arithmetic
One attempt is the weighted sum of the 3 outcomes. Of 100 attempts, 65 reach a machine and 3 reach a person. Thus the carrier bills 68 answered calls, and OpenAI reads 68 calls:
- Machine calls: 0.65 × $0.042 = $0.027
- Person calls: 0.03 × $0.076 = $0.002
- Carrier minutes: 0.68 × $0.009 = $0.006
- Post-call analysis: 0.68 × $0.003 = $0.002
- One attempt = $0.038 at the measured 16 s per machine call.
- One attempt = $0.026 at 9 s per machine call, the target on What a call costs ($0.027 if Retell bills its 10 s minimum).
One lead = attempts per lead × 1 attempt. That is 12 attempts on average and 15 at the maximum.
The machine calls are most of the cost. 65 of 100 attempts reach a machine. Each machine call costs 16 s of the voice service. People cost almost nothing in total, because so few answer. An answered call with a person costs about $0.088 all in: $0.076 for Retell, $0.009 for Telnyx and $0.003 for OpenAI.
10,000, 50,000 and 100,000 leads
Campaign length: 10,000 leads run over 30 business days. 50,000 and 100,000 leads run over 60 business days. The rent covers that length. The machine call is at the measured 16 s.
Average case: 12 attempts per lead, $0.038 per attempt
| Leads | Calls | Retell | Telnyx minutes | OpenAI | Numbers to buy | Rent | Total | Per lead |
|---|---|---|---|---|---|---|---|---|
| 10,000 | 120,000 | $3,540 | $730 | $240 | 67 | $100 | $4,620 | $0.46 |
| 50,000 | 600,000 | $17,700 | $3,670 | $1,220 | 167 | $500 | $23,090 | $0.46 |
| 100,000 | 1,200,000 | $35,390 | $7,340 | $2,450 | 334 | $1,000 | $46,180 | $0.46 |
Maximum case: 15 attempts per lead, no lead ever answers, $0.035 per attempt
| Leads | Calls | Retell | Telnyx minutes | OpenAI | Numbers to buy | Rent | Total | Per lead |
|---|---|---|---|---|---|---|---|---|
| 10,000 | 150,000 | $4,080 | $880 | $290 | 84 | $130 | $5,380 | $0.54 |
| 50,000 | 750,000 | $20,410 | $4,390 | $1,460 | 209 | $630 | $26,890 | $0.54 |
| 100,000 | 1,500,000 | $40,820 | $8,780 | $2,930 | 418 | $1,250 | $53,770 | $0.54 |
The machine hang-up time changes the total most. The target on What a call costs is 9 s, not 16 s. At 9 s, 100,000 leads cost $31,900 in the average case. That is $0.32 per lead on average, and $0.36 per lead at the maximum.
How to read the tables:
- Retell is 77% of the cost. Telnyx minutes are 16%. OpenAI is 5%. Number rent is 2%.
- The per-lead figure does not change with the client's size. It changes with 3 inputs: attempts per lead, seconds per machine call, and the Retell rate per minute.
- The count of phone numbers depends on speed, not on size: 100,000 leads over 120 business days need 167 phone numbers, not 334, for the same total.
What makes the cost lower, in order of size
| Change | Effect |
|---|---|
| Hang up on a machine in 9 s, not 16 s | −32% (−27% if Retell bills its 10 s minimum) |
| Fewer attempts per lead (a shorter cadence, or drop a lead after a set number of voicemails) | In proportion: 6 attempts = half the cost |
| Shorten the agent's prompt below 4,000 tokens, so Retell drops the long-prompt surcharge (19.2% of the Retell cost) | About −15% |
| Keep the production agent free of the other staging add-ons (personal-data scrubbing, colloquial mode, guardrail) | Already the case. The staging rate of $0.30 per minute is about +90% on the Retell rate. |
| A cheaper AI model ($0.013 per minute instead of $0.045) | About −15% (−20% of the Retell cost) |
What the model leaves out
The tables cover the calls and the number rent only. They leave out:
- Text messages, at about $0.011 each.
- Number reputation lookups.
- Hosting, about €179 a month for the whole platform.
- The do-not-call list subscription, $449 a month.
- The email service.
- The transfer, and the people who handle it.
The platform's spend ledger also does not record number rent, number reputation lookups, hosting, the do-not-call list subscription or the email service. The daily spend ceiling therefore does not count them either.
What is still an estimate
Two inputs are a model, not a measurement:
- The attempts per lead (approximately 12 on average).
- The outcome mix of 65 machine, 3 person and 32 no answer per 100 calls.
Production has run only the QA persona campaign: 6 personas, 7,220 attempts. 93% of the attempts did not connect, because the personas were not available. The first real campaign gives the measured mix. Engineering gets the mix from the stored call outcomes, grouped by who answered.
The count of phone numbers depends on speed, not on size
The count of phone numbers to buy follows the dials per day, not the total of leads. The arithmetic has 4 lines:
- Dials per day = leads released per day × attempts per lead.
- Working numbers = dials per day ÷ 75, rounded up.
- Numbers to buy = working numbers × 1.25.
- Rent = numbers × $1.00 × months.
The 334 in the average table comes from this arithmetic. 100,000 leads over 60 business days is 1,667 leads a day. At 12 attempts, that is 20,000 dials a day. Divided by 75, that is 267 working phone numbers. With the 25% reserve, that is 334 phone numbers to buy. The same 100,000 leads over 120 business days halve the dials per day, so 167 phone numbers are enough.
Two more facts shape the plan:
- Warm-up. A new phone number allows 25 dials a day on days 1–5 and 50 dials a day on days 6–10. Thus a new pool delivers 1/3 of its capacity in week 1 and 2/3 in week 2. The page on bringing a number into service explains why.
- Calls at the same time. 20,000 dials over an 8-hour day is approximately 0.7 dials a second. A dial takes 30–40 s from ring to hang-up. Thus this client has approximately 25 calls open at the same time. The account limit of the carrier must cover the sum over all clients. Billing each client explains the limits and how to raise them.
Every input here is in the code or in the cadence configuration. The design of a "Buy numbers" screen exists. The screen takes leads and days, shows this table and places the order. LeadRefinery has deferred the screen. You buy phone numbers by hand for now.