Every comparison of AI voice agent pricing answers the same question: which platform charges least per minute. That question matters until your volume gets large enough, at which point a different question matters more.
At what call volume does building your own stack cost less than renting someone else’s?
This guide answers that with the actual numbers, including the build cost most comparisons leave out entirely.
What a Hosted Platform Actually Costs
Published rates are lower than real rates, and the gap is consistent enough to plan around.
Platforms advertise a headline figure covering part of a call. Vapi publishes $0.05 a minute for hosting and passes speech recognition, the model, the voice, and telephony through at provider cost. Retell publishes a range from roughly $0.07 to $0.31 depending on which components you select. Bland bundles its stack at about $0.14 a minute with no platform fee, or $0.12 with a monthly commitment. Synthflow and PolyAI publish no per-minute rate for a managed agent at all, with Synthflow enterprise agreements reported starting around $30,000 a year.
Stack the layers and a production deployment on a hosted platform generally lands between $0.11 and $0.25 a minute, with most published analyses converging near $0.12 to $0.18.
Cost layer | Typical range per minute |
Speech-to-text | $0.0015 to $0.024 |
Language model | $0.003 to $0.08 |
Text-to-speech | $0.02 to $0.10 |
Telephony | $0.005 to $0.014 |
Platform orchestration fee | $0.05 to $0.14 |
All-in on a hosted platform | $0.11 to $0.25 |
Rates reflect published pricing as of late 2026 across several independent comparisons. Confirm current figures directly with any vendor before modeling, since this category reprices frequently.
What a Custom Stack Costs to Run
Remove the platform orchestration fee and the picture changes substantially, because that fee is the single largest line in the table above.
A custom stack assembles the same components directly: a speech-to-text provider, a model, a voice, and a carrier. Published analyses put the usage cost of a custom-assembled stack at roughly $0.04 a minute, and component pricing supports that. Deepgram and AssemblyAI streaming transcription sit in fractions of a cent. Telnyx publishes US outbound around half a cent a minute against Twilio’s roughly 1.4 cents. Model and voice selection drive most of the remaining variance.
The saving is real. It is also not free, because somebody has to build and operate the thing.
The Build Cost Nobody Puts in the Comparison
A custom voice agent build is commonly quoted in the range of $25,000 to $120,000 depending on scope, integration depth, and how much the system has to handle beyond a single straightforward call flow.
That range is wide for a reason. A single-purpose inbound agent with one integration sits near the bottom. A multi-intent agent connected to a CRM, a scheduling system, and a billing platform, with escalation logic and compliance requirements, sits near the top.
Hosting and operations are separate and ongoing. A self-hosted stack needs infrastructure, monitoring, on-call coverage, and maintenance as providers change their APIs underneath you.
The Break-Even Math
Take the middle of each range. A hosted platform at $0.15 a minute against a custom stack at $0.04 a minute saves $0.11 per minute.
Against a $60,000 build, break-even arrives at roughly 545,000 minutes. Against a $25,000 build, roughly 227,000 minutes. Against $120,000, roughly 1.1 million minutes.
Build cost | Minutes to break even at $0.11 saved | Roughly, at 50,000 min/month |
$25,000 | ~227,000 | ~4.5 months |
$60,000 | ~545,000 | ~11 months |
$120,000 | ~1,090,000 | ~22 months |
Those figures exclude hosting and maintenance, which push every break-even point further out. Treat them as the optimistic case, not the expected one.
The practical read: below roughly 20,000 minutes a month, a hosted platform almost always wins on total cost. Above roughly 100,000 minutes a month, a custom build usually pays back inside a year. Between those, the answer depends on your specific build scope and how much engineering capacity you already have.
When Cost Is Not the Deciding Factor
Several situations make a custom build correct regardless of where break-even lands.
Data residency or compliance requirements that a managed platform cannot satisfy remove the choice entirely. So does integration depth beyond what a platform’s connectors anticipated, which is a frequent reason teams reach the limits of a hosted product well before they reach the limits of its pricing.
Latency control matters in the other direction. A platform abstracts the pipeline, which is convenient until you need to tune the specific stages that determine whether a conversation feels natural.
Equally, several situations make a platform correct regardless of volume. If you have no engineering capacity to operate infrastructure, a custom build converts a predictable per-minute cost into an unpredictable operational burden. That trade is rarely worth it.
How to Run This Calculation for Your Own Numbers
Start with your real monthly minutes, not your projected ones. Most volume forecasts in this category are optimistic, and break-even math is unforgiving of a forecast that lands at half of plan.
Get a scoped build quote rather than using a range. The $25,000 to $120,000 spread is too wide to model against, and your actual number depends on integration count more than anything else.
Add hosting and maintenance as a running line, not a one-time cost. A custom stack that saves $0.11 a minute and costs $4,000 a month to operate has a meaningfully different break-even than one that costs $800.
Then compare against a real platform quote at your volume, since most platforms discount materially above a certain commitment and the published rate is not what you would actually pay.
Where RTC LEAGUE Fits
We build the custom side, which means we have an obvious interest in you choosing it. So the useful thing to say is when not to.
If you are running under about 20,000 minutes a month with no compliance constraint and no integration a platform cannot handle, buy the platform. You will spend less and move faster, and we would be a worse answer than Vapi or Retell.
Where a custom build earns its cost is volume above roughly 100,000 minutes a month, a compliance or residency requirement that rules out managed infrastructure, or integration depth that a connector was never going to reach. That is the work our AI voice platform and agentic AI development practice covers, usually with CRM integration attached because the integration depth is frequently the actual reason the platform stopped working.
If you want the break-even run against your real numbers rather than the ranges in this article, that is a short conversation and we will tell you if the answer is to stay on a platform.






