TL;DR & Quick Summary
Most published pricing for AI voice agents tells you what a vendor charges. This teardown covers what the thing costs to operate — every component, priced per minute of call time.
The short version, for a three-minute inbound call:
- Telephony: ~$0.009/min — effectively fixed
- Speech-to-text: ~$0.008/min — effectively fixed
- LLM inference: ~$0.007–$0.038/min — swings ~5x on model choice
- Text-to-speech: ~$0.020–$0.070/min — usually the largest single line
- Orchestration and hosting: ~$0.004–$0.050/min — depends on self-hosted vs platform
All-in: roughly $0.04 to $0.13 per minute.
- Key Takeaway: The language model is not the expensive part. Voice synthesis usually is — and the engineering to build and maintain the agent costs far more than every per-minute line combined. Build for integration, not to save on minutes.
- Get Started: Want the cost model run against your actual call volume and booking stack? Schedule a Strategy Call with Cogniq AI or explore our voice AI automation services.
About the rates in this teardown. The figures below are representative published rates from mainstream providers — Twilio, Deepgram, ElevenLabs, Cartesia, Vapi, Retell — at the time of writing. Component pricing in this market moves quickly and volume tiers change the arithmetic. Treat every number here as a model you re-run with current rates, not a quote. The structure of the model is what stays true; the inputs will not.
Why Nobody Publishes This
Per-minute component economics are the basis of every AI receptionist vendor's margin. Publishing them invites exactly the question vendors would rather you did not ask: what am I actually paying for?
The answer, for a well-run vendor, is legitimate — engineering, integrations, uptime, support, and the person who answers when the agent mishandles a call at 2am. Those are real costs and they deserve real margin. But you cannot evaluate whether a price is fair without knowing the floor beneath it.
This teardown gives you that floor. For the other side of the comparison — what vendors actually charge, normalised across nine providers — see our AI receptionist pricing comparison. This post is the cost side; that one is the price side.
The Five Cost Lines
Every voice agent, regardless of who builds it, runs the same loop: audio comes in over a phone network, gets transcribed, gets reasoned about, gets spoken back. Four services and some orchestration.
| Component | What It Does | Typical Rate | Per Minute of Call |
|---|---|---|---|
| Telephony | Carries the call over the PSTN | ~$0.0085/min inbound, plus ~$1.15/month per number | ~$0.009 |
| Speech-to-text | Streams caller audio to text | ~$0.0077/min streaming | ~$0.008 |
| LLM inference | Decides what to say and which tools to call | Per token, varies by model | ~$0.007–$0.038 |
| Text-to-speech | Renders the reply as speech | Per character, varies sharply by provider | ~$0.020–$0.070 |
| Orchestration | Turn-taking, barge-in, tool routing, state | Platform fee or self-hosted compute | ~$0.004–$0.050 |
Two of these are effectively fixed. Two of them are where your bill is actually decided.
Line 1 and 2: Telephony and Speech-to-Text
Take these as constants and stop optimising them.
Telephony runs around $0.0085 per inbound minute on Twilio for a US local number, plus roughly $1.15 per month to rent the number itself. Toll-free costs more. International inbound varies widely by country, and if you serve callers across the EU or APAC you will want per-region numbers, which multiplies the fixed line but not the per-minute one.
Streaming speech-to-text on Deepgram's Nova models lands near $0.0077 per minute. Competitors cluster within a cent of that. Batch transcription is cheaper, but batch is useless here — a voice agent needs partial transcripts within a few hundred milliseconds to interrupt naturally.
Combined, these two lines total roughly $0.017 per minute and are nearly the same whichever vendor or stack you choose. At 600 minutes a month that is about $10. Any effort spent shaving these is misdirected.
Line 3: LLM Inference — Smaller Than You Think
This is the line teams over-estimate most, and it is worth doing the arithmetic properly.
A three-minute booking call is roughly 12 to 16 conversational turns. On each turn you resend the system prompt, the tool definitions, and the conversation so far — so token consumption grows through the call rather than staying flat.
A realistic profile for a well-built agent:
- System prompt and tool schemas: ~1,200 tokens, resent every turn
- Average total context across the call: ~2,500 tokens per turn
- Fourteen turns: ~35,000 input tokens cumulative
- Replies are short by design: ~40 tokens per turn, so ~560 output tokens total
Now price that two ways:
| Model Tier | Input Rate | Output Rate | Cost per 3-min Call | Per Minute |
|---|---|---|---|---|
| Fast small model | ~$0.60/M | ~$2.40/M | ~$0.022 | ~$0.007 |
| Frontier model | ~$3.00/M | ~$15.00/M | ~$0.113 | ~$0.038 |
Model choice moves this line by roughly 5x — and yet even the expensive option is a minority of total cost. Two practical consequences follow.
First, prompt caching matters more than model price. That 1,200-token system prompt is identical on every turn. Caching it cuts input cost substantially on providers that support it, and the saving compounds with call length.
Second, latency should drive model selection, not price. In voice, a 900ms pause reads as a broken call. Callers talk over it, the agent restarts, and the interaction degrades. The cheaper small model is frequently the better choice on quality grounds because it responds faster — and it happens to cost less. Where a frontier model earns its place is complex reasoning mid-call: multi-constraint scheduling, eligibility logic, genuine ambiguity. A sensible design routes most turns to a fast model and escalates only the turns that need it.
Line 4: Text-to-Speech — Usually Your Largest Line
Here is the finding that surprises most teams building their first agent.
In a three-minute call the agent speaks perhaps 55% of the time — roughly 1.6 minutes, about 240 words, or ~1,400 characters of synthesised speech.
| TTS Tier | Rate per 1,000 chars | Cost per 3-min Call | Per Minute |
|---|---|---|---|
| Premium voice | ~$0.090 | ~$0.126 | ~$0.042 |
| Mid-tier voice | ~$0.050 | ~$0.070 | ~$0.023 |
| Budget / efficient voice | ~$0.030 | ~$0.042 | ~$0.014 |
At the premium tier, text-to-speech costs more than telephony, speech-to-text, and a fast LLM combined. It is the single largest variable in the stack, and it is the first place to look when a per-minute number comes back higher than expected.
This is a genuine product decision, not a pure cost one. A noticeably synthetic voice increases hang-ups, and a hang-up costs you the booking — which is worth far more than the fraction of a cent you saved. But the gap between premium and good has narrowed considerably, and providers like Cartesia and Deepgram's Aura line now deliver quality that most callers do not consciously register as synthetic. Testing a cheaper voice against your own callers is usually the highest-return experiment available.
Two ways to cut this line without touching voice quality:
- Cache your fixed phrases. Greetings, hold messages, confirmations, and closings are identical on every call. Render them once, store the audio, replay it. In a typical booking flow this removes 20–30% of synthesised characters permanently.
- Write shorter replies. Verbose agent turns cost money twice — once in synthesis, and again in the caller's patience. Tight replies are both cheaper and better.
Line 5: Orchestration and Hosting
Something has to manage turn-taking, barge-in detection, tool calls, retries, and conversation state. Two routes:
Platform-managed — Vapi and Retell handle orchestration and pass component costs through. Vapi has published platform hosting at $0.05 per minute with model costs at cost; Retell publishes an all-in band of roughly $0.07 to $0.31 per minute. You are buying working turn-taking, which is genuinely difficult engineering, and you are paying a per-minute rate forever.
Self-hosted — a modest always-on container handling websocket audio streams plus monitoring and logging. Call it $20–$40 per month in compute for low-to-moderate concurrency. Across 600 minutes that is about $0.004 per minute — but it only stays cheap if you have already solved barge-in, reconnection, and state handling, which is engineering time rather than infrastructure spend.
The honest read: platform orchestration is worth paying for until volume makes it painful. It is the component with the worst effort-to-value ratio to rebuild yourself.
Putting It Together: A Real Monthly Bill
Take the profile from our vendor pricing comparison so the two posts are directly comparable — 200 calls per month at three minutes each, 600 total minutes.
| Line | Rate/min | Monthly at 600 min |
|---|---|---|
| Telephony | $0.009 | $5.40 |
| Speech-to-text | $0.008 | $4.80 |
| LLM (fast model, cached prompt) | $0.007 | $4.20 |
| Text-to-speech (mid-tier) | $0.023 | $13.80 |
| Orchestration (self-hosted) | $0.004 | $2.40 |
| Usage subtotal | $0.051 | $30.60 |
| Phone number rental | — | $1.15 |
| Monitoring, logging, error tracking | — | ~$25.00 |
| Total running cost | — | ~$57/month |
Against that, mainstream AI-only vendors at the same volume land between $129 and $399 per month.
So a vendor's gross margin on usage sits somewhere around 55% to 85%. That is not a scandal — it funds the engineering, the integrations, the support, and the uptime you are not building yourself. But it does tell you the size of the prize if you build, which sets up the only question that matters.
The Cost That Dwarfs All of This
Everything above is rounding error next to engineering.
A production voice agent — not a demo — needs prompt and flow design, integration with your booking system and CRM, tool functions that write reliably into real records, failure handling for when a tool call times out mid-call, transfer-to-human logic, call logging, and testing across accents, background noise, and interruptions.
Realistically that is 60 to 140 hours for a first production build. At a stated blended rate of £75/hour, that is £4,500 to £10,500. Maintenance runs a further 4 to 8 hours per month — prompt tuning, integration drift, handling the edge cases real callers surface — or roughly £300 to £600 monthly.
(These are illustrative figures at a stated rate, not benchmark data. Your scope and rate will differ.)
Now the break-even, done honestly:
- Vendor at this volume: ~$200/month
- Self-built running cost: ~$57/month
- Monthly saving: ~$143
- Build cost at the low end: £4,500 ≈ $5,700
- Payback on cost alone: about 40 months — before counting a single hour of maintenance.
Add maintenance and the saving inverts: you are now spending more than the vendor charged.
At typical service-business volume, building to save money on minutes is a losing trade. Anyone who tells you otherwise is either quoting you a demo rather than a production system, or has not counted the maintenance.
When Building Actually Wins
Building is the right call when the driver is capability, not price:
- Integration vendors cannot do. Writing into an EHR, a legacy booking database, a custom CRM, or a proprietary dispatch system. This is the most common legitimate reason, and it is not a cost argument at all — the vendor simply cannot do the job. Our AI automation audit playbook covers how to establish whether this applies to you before committing.
- Volume above roughly 3,000–5,000 minutes per month. Per-minute margin compounds. At 5,000 minutes the same margin gap is over $1,000 monthly, and payback moves inside a year.
- Compliance requiring architectural control. Data residency, a signed BAA, retention limits, or audited access logging. Platform add-ons for HIPAA and zero data retention have been published at $2,000 and $1,000 per month — at which point control and cost point the same direction. Our LLM data security framework sets out what to verify.
- The agent must complete work, not just answer. Checking inventory, applying eligibility rules, processing a payment, updating multiple systems in one flow. Answer-and-route products are priced and built for a narrower job.
- Ownership. The agent is an operational asset. Vendor pricing changes, products sunset, terms move.
If none of these apply, buy the subscription. That is a genuine recommendation, and it is the one we give clients when it is true.
The Costs Everyone Forgets
Budgets built from the table above still come in low, because these lines get missed:
- Failed and abandoned calls bill anyway. A caller who hangs up at fifteen seconds still cost telephony, streaming STT, and a synthesised greeting. Wrong numbers and spam calls bill too. Budget against total inbound minutes, not successful outcomes.
- Barge-in doubles some processing. Natural interruption handling means running STT continuously while the agent speaks. That is deliberate — it is what makes the agent feel human — but it is not free.
- Testing costs real minutes. A serious pre-launch test pass is several hundred minutes of telephony, STT, and TTS across accents and noise conditions. Budget it explicitly.
- Recording storage and retention. Call recordings and transcripts accumulate, and in regulated contexts retention is mandatory rather than optional.
- Concurrency, not just volume. Ten simultaneous callers at peak requires provisioning for ten, whatever your monthly average says. A clinic taking 80% of its calls in a two-hour window prices very differently from one with flat distribution.
- Per-region numbers. Multi-location businesses rent a number per site. The per-minute cost is unchanged; the fixed line scales linearly.
How to Run This Model Yourself
Before requesting a single quote, get two numbers from your phone provider's call log — not from memory, which reliably underestimates:
- Total inbound minutes per month, including abandoned calls.
- Peak concurrent calls, which sets provisioning.
Then work through it in order:
- Multiply total minutes by $0.017 for the fixed telephony and STT floor.
- Add $0.007–$0.038 per minute for the LLM, depending on whether your flow genuinely needs frontier reasoning.
- Add $0.014–$0.042 per minute for TTS, and test a mid-tier voice against your callers before assuming you need premium.
- Add orchestration: a platform per-minute rate, or $20–$40 monthly self-hosted.
- Add fixed costs: number rental, monitoring, storage.
- Then add engineering — build hours amortised over a realistic lifespan, plus monthly maintenance. This is the line that decides build versus buy, and it is the line most models omit.
Compare that total against vendor pricing at your actual volume. If the gap does not clear the engineering cost inside 18 months, buy — and revisit when volume or integration requirements change. For the wider question of what returns these systems actually produce, our AI agent ROI benchmarks covers what the primary data supports.
Conclusion
The per-minute economics of AI voice agents are not mysterious, and they are not where most of the money is. Telephony and speech-to-text are fixed and trivial. The language model is cheaper than almost everyone expects. Text-to-speech is quietly the largest usage line and the best place to optimise. Orchestration is worth renting until volume says otherwise.
And all of it is dominated by engineering — which is precisely why "build it ourselves and save on the subscription" is usually the wrong reason to build, and why "no vendor can write into our booking system" is usually the right one.
If you want this model run against your real call volume, your real booking stack, and an honest answer about whether you should build at all, that is a conversation worth having — including when the answer is that a subscription serves you better.
Schedule a Strategy Call with Cogniq AI and we will run the numbers with you, or read more about how we approach voice AI automation.
