If you've priced an AI voice agent in 2026, you've seen the number: $0.07 a minute. Sometimes as low as $0.05. It looks almost free next to a human calling your leads.
Then the first invoice shows up and it's two to four times higher.
Here's the honest version, from someone who builds these agents for clients: the real all-in cost of an AI voice agent is $0.13–$0.31 per minute, not the $0.05–$0.09 on the pricing page. The headline rate is the platform's orchestration fee. The speech-to-text, the language model, the text-to-speech, and the phone line all bill separately, and stacked together they roughly double or triple the number you were quoted. None of that makes voice agents a bad deal — it makes them a deal you need to model honestly before you sign up.
Why the Sticker Price Is Almost Never the Real Price
Every voice AI platform — Retell, Vapi, Bland — has to run four moving parts on every call:
- Speech-to-text (STT) turns what the caller says into text.
- The language model (LLM) decides what to say back.
- Text-to-speech (TTS) turns that reply into a voice.
- Telephony actually carries the call over the phone network.
When a pricing page says "$0.07/minute," that's usually just the platform's fee for gluing those four together and managing the conversation state. The four components underneath get billed through at usage rates — and one of them, text-to-speech, is frequently the most expensive line on the whole invoice.
That's the gap that surprises people. You're not being lied to; you're reading the orchestration fee and assuming it's the total.
The Real Per-Minute Breakdown
Here's what each layer actually costs, based on 2026 provider rates and the stacks we run in production:
| Component | Real cost per minute | Notes | |-----------|---------------------|-------| | Speech-to-text (STT) | $0.005–$0.02 | Deepgram Nova streaming is ~$0.0077/min | | Language model (LLM) | $0.02–$0.10 | GPT-4o-class inference at conversational token volume | | Text-to-speech (TTS) | $0.05–$0.18 | Usually the single most expensive component | | Telephony (US) | $0.005–$0.03 | Twilio adds ~$0.02/min per call | | Platform orchestration | $0.05–$0.14 | The number on the pricing page | | All-in total | $0.13–$0.31 | What you actually pay |
A few things jump out once you see it laid out:
Text-to-speech is where the money goes. Premium voices from providers like ElevenLabs can run $0.05–$0.18/minute on their own. Switching to a cheaper engine like Deepgram Aura can cut $0.04–$0.08 off every minute — a 30–50% reduction in total call cost on some configurations. The voice you pick is a bigger cost lever than the model you pick.
Telephony is bigger than it looks. At ~$0.02/minute, Twilio alone can be 35–50% of total cost on the leanest stacks. It's easy to forget the phone line is a metered cost too.
Premium and cloned voices stack on top. A custom cloned voice can add another $0.05–$0.20/minute. Great for brand, brutal on unit economics at volume.
Retell vs Vapi vs Bland: The All-In Ranges
Platform choice moves the number less than most people expect, because they all run the same four-layer stack underneath. Here's where the three most common platforms land once you add everything up:
| Platform | Advertised rate | Real all-in | Best for | |----------|----------------|-------------|----------| | Retell AI | ~$0.07/min | $0.13–$0.31/min | Low-code teams that want telephony, warm transfer, and compliance built in | | Vapi | $0.05–$0.10/min platform fee | $0.10–$0.30/min | Developers who want to hand-pick their own LLM, voice, and telephony | | Bland AI | ~$0.09 flat | $0.09–$0.20/min | Simpler, more bundled flows where you don't need to tune the stack |
The takeaway: you're choosing an architecture, not a discount. Retell is low-code and ships warm transfer, native SIP, and HIPAA/SOC 2 across plans, which is why we default to it for client builds. Vapi gives developers full control of every component but leaves telephony features like warm transfer for you to assemble. Bland trades flexibility for a simpler bundled experience. The all-in per-minute cost between them is close enough that ease of building and maintaining the agent — not the sticker rate — should drive the decision. We covered why we standardize on Retell in our agency guide to Retell voice agents.
What the Cheapest Possible Stack Really Costs
If you build it yourself from raw components — Groq running Llama for the LLM (~$0.001/min), Deepgram Nova for STT (~$0.008/min), Deepgram Aura for TTS (~$0.036/min), and a direct SIP bridge instead of Twilio (~$0.002/min) — you land around $0.05/minute in pure component cost.
That sounds like a win until you remember what that number leaves out: orchestration, error handling, retries, compliance recording, monitoring, logging, and the engineering hours to build and maintain all of it. The raw component cost is the floor, not the price. For most agencies and service businesses, hand-rolling the cheapest stack to save eight cents a minute is a false economy — you spend the savings ten times over in engineering and babysitting. This is the same trade-off we walk clients through when they ask whether to migrate their automations off Zapier and Make to self-hosted n8n: cheaper per-unit, more to own.
The Hidden Fees Nobody Puts on the Pricing Page
Model these before you commit, because they quietly add 15–30% to any per-minute estimate:
- Silence billing. Some platforms charge for dead air, hold time, and IVR navigation, not just active conversation. On long calls with transfers, this adds up fast.
- Concurrency limits. The cheap tier often caps how many calls run at once. Hit your Monday-morning rush and you're forced onto a higher plan.
- Compliance surcharges. HIPAA-eligible configurations frequently add $1,000/month or more before you've made a single call.
- Toll-free and international premiums. A toll-free number and international routing both cost more per minute than a standard local line.
- Recording storage. If you record calls for QA or compliance, that storage is a recurring line item too.
None of these are dealbreakers. They're just the difference between a budget that holds and one that blows out in month two.
The Only Cost Number That Actually Matters
Per-minute pricing is the wrong altitude to make the decision at. The number that matters is cost per outcome — cost per booked appointment or per qualified lead.
Run the math the way we do for clients. Take the worst-case all-in rate, $0.31/minute. A 4-minute qualification call costs about $1.24. If that agent books one discovery call for every twelve conversations, your cost per booked call is under $15 — running 24/7, answering the 11pm Saturday lead your team will never call back in time.
Compare that to a human SDR. Loaded with salary, tools, and downtime, a person books meetings at a cost per booked call in the hundreds of dollars, and only during business hours. For high-volume, repetitive calling — lead qualification, appointment reminders, first-touch follow-up — the voice agent wins on cost by an order of magnitude, even at the expensive end of the per-minute range. We broke down one of these builds end to end in how we cut lead response time to five minutes with Retell.
The place it flips is low-volume, high-complexity, high-value calls, where a human's judgment closes materially more revenue per conversation. That's not most of the calls a service business makes. Most calls are repetitive — which is exactly why a voice agent is worth the honest math.
How to Price Your Own Agent Before You Commit
- Split the quote. Treat the advertised rate as the orchestration fee only. Add STT, LLM, and TTS as separate line items.
- Add telephony. ~$0.02/minute for US calls, more for toll-free or international.
- Add 15–30% for hidden fees. Silence billing, concurrency, compliance, premium voices.
- Use your real average call length, not the vendor's tidy example.
- Divide by booked outcomes. Cost per appointment is the number that decides it.
Do that and you'll walk in knowing your true unit economics instead of discovering them on invoice day.
Get the Real Numbers for Your Use Case
The ranges above are honest, but your actual cost depends on your call volume, average handle time, the voice you want, and how much of the stack you own. We build and run voice agents for clients across lead qualification, booking, and follow-up — and we'll model the true per-minute and per-outcome cost for your specific case before you spend a dollar on a platform.
Get a free automation audit and we'll show you exactly what a voice agent would cost — and what it would save — for your business.
Sources: provider pricing from Retell AI, Deepgram, and Twilio as of mid-2026. Component ranges reflect stacks we run in production; your mileage will vary with voice, model, and call length.