Every AI phone platform advertises a per-minute price, and almost none of them mean the same thing. Vapi's $0.05 a minute is its platform fee before you've paid for a single word of transcription. Retell's $0.07 headline rate assumes a budget model and a basic voice. The number that actually matters is the all-in cost per minute, multiplied by the minutes you'll really handle, plus the fees that only surface once you need more lines or a BAA.
This article breaks down both platforms' current pricing line by line, then puts a self-hosted voice agent on AWS next to them at five call volumes. Prices were checked against the Vapi, Retell, Twilio, Deepgram, ElevenLabs, and AWS pricing pages on October 22, 2026. They change often, so treat this as a snapshot.
The Quick Verdict
Managed platforms are cheaper until you're handling several thousand minutes a month, and much faster to start. Self-hosting wins on cost at sustained volume, and on control when call data cannot leave your account.
Under ~7,000 min/month
A fixed GPU bill can't compete with pay-per-minute at low volume, and you skip building the stack.
Frontier models and premium voices
A single self-hosted GPU won't match GPT-5.5 or ElevenLabs quality.
50,000+ min/month
The managed bill grows with every minute; yours grows with peak concurrency, which is much slower.
HIPAA or strict data rules
Vapi's HIPAA add-on alone is $2,000/month; Retell lists a self-signed BAA with no separate fee, which changes the break-even before volume does.
What a Minute of AI Phone Call Is Made Of
Regardless of platform, every minute pays for the same five layers:
| Layer | What It Does | Typical Managed Price |
|---|---|---|
| Platform or voice engine | Orchestration, turn-taking, interruptions | $0.05 (Vapi), $0.055 (Retell) |
| Speech-to-text | Turns caller speech into text | ~$0.01 |
| LLM | Decides responses and which tools to call | $0.0016 to $0.38, by model |
| Text-to-speech | The agent's voice | $0.015 to $0.10, by voice |
| Telephony | Phone number and the call itself | $0 to $0.015 |
The LLM and the voice are the swing factors. The platform fee is fixed per minute. The model and voice you pick can more than double the total, so any quote that doesn't name both isn't a quote.
Vapi Pricing: Platform Fee Plus Pass-Through
Vapi charges a hosting fee and passes every provider through at cost. You bring the models or pick from its list, and pay each one separately.
| Item | Price |
|---|---|
| Vapi hosting | $0.05/min, prepaid |
| Transcriber (Deepgram Nova-3) | $0.0095-$0.0099/min |
| LLM (OpenAI) | $0.008-$0.045/min |
| Voice (ElevenLabs API) | $0.015-$0.024/min |
| Telephony | Vapi telephony, SIP, and WebRTC free; Twilio $0.0085 in / $0.014 out; Telnyx $0.0055 |
| Concurrency | 4 lines on the free tier, 10 on Core ($29/month); extra lines $10/month each |
| HIPAA | $2,000/month add-on |
| Zero Data Retention | Included with any paid support package |
Vapi's own example puts 1,000 minutes at $82-$129, which works out to $0.082-$0.129/min all-in on free telephony. Bring your own Twilio number and add $0.0085/min for inbound calls. The key caveat: HIPAA is a flat $2,000/month add-on, more than two self-hosted GPU stacks before a single call is made. Zero Data Retention needs a paid support package, which starts at $29/month.
Retell Pricing: Bundled Engine, Metered Components
Retell bundles transcription into its voice engine, then charges separately for the LLM, voice, and phone line. The published all-in range is $0.07-$0.31/min, and the spread comes almost entirely from the LLM you choose.
| Item | Price |
|---|---|
| Retell voice engine | $0.055/min |
| LLM (standard tier) | $0.0016 (GPT-5 nano) to $0.16 (GPT-5.5); $0.345-$0.38 (Realtime) |
| Voice (TTS) | $0.015/min for Retell, Minimax, Fish, Cartesia, OpenAI; $0.04-$0.10/min for ElevenLabs |
| Telephony | $0.015/min through Retell; free on your own SIP trunk |
| Phone numbers | $2/month |
| Concurrency | 20 calls included; $8/month per extra line |
| HIPAA / BAA | Self-signed BAA on pay-as-you-go; no separate fee listed |
The example stack on Retell's pricing page (a $0.04 LLM, Retell's $0.015 voice, and $0.055 infrastructure) lands at $0.11/min before telephony. With Retell's phone line at $0.015, that becomes $0.125/min. Swap in Claude 4.5 Haiku at $0.025 and bring your own SIP trunk, and it drops closer to $0.095. Go the other way with GPT-5.5 at $0.16 and ElevenLabs voice at $0.04, and you're at $0.255 before telephony.
The compliance picture is where Retell diverges sharply from Vapi: a self-signed BAA is available on pay-as-you-go, and the pricing page lists no separate fee for it. Enterprise adds SSO, RBAC, a dedicated server, and custom MSA/DPA/BAA terms behind custom pricing, but the basic legal path for a clinic doesn't need an enterprise contract.
Self-Hosted Phone Agent on AWS
Self-hosted, the per-minute cost nearly disappears and a fixed monthly bill takes its place. The stack: LiveKit for transport, faster-whisper for speech-to-text, an open-weights LLM on vLLM, and Kokoro for the voice.
GPU instance
A g6.xlarge (NVIDIA L4, 24 GB) at $0.8048/hour on-demand runs speech-to-text and an 8B-class model. That's $587.50/month around the clock, or about $269.66/month on a three-year reserved commitment. Spot pricing drops as low as $0.4013/hour (~$293/month), though interruption rates exceed 20%.
CPU instance
LiveKit, the agent workers, Kokoro, and the turn detector run on CPU. Budget roughly $130/month for a 4-vCPU instance. Keep it in the same region as the GPU.
Telephony
You still need a carrier. Twilio charges $0.0085/min for inbound calls on a local number, $0.014/min for outbound, and $1.15/month per number. Telnyx is cheaper: Vapi passes its calls through at $0.0055/min, against $0.008 for Twilio.
What one GPU handles
A 24 GB GPU is pilot-sized: a handful of concurrent calls. Faster-whisper serializes inference calls at the GPU kernel level, so peak throughput is limited by GPU saturation, not thread count. How many concurrent calls you actually get depends on your model, prompt length, and latency target, so load-test before sizing. Past that, you add GPUs for peak concurrency, not total minutes.
That last point is why self-hosting scales so well on cost. Managed platforms bill every minute. A GPU bills every hour whether it's busy or idle, so what you pay for is your busiest hour, not your total volume.
Monthly Cost at Five Call Volumes
Here's each option at five monthly volumes. The managed columns use each platform's own published estimates. The self-hosted column assumes calls arrive during business hours (8 hours/day, 22 days/month), peak at 1.5× the average, and 5 concurrent calls per GPU. That last figure is an assumption for illustration; test your own workload.
| Minutes/month | Self-hosted on AWS | Vapi ($0.082-$0.129) | Retell ($0.07-$0.31) |
|---|---|---|---|
| 1,000 | ~$727 (1 GPU) | $82-$129 | $70-$310 |
| 10,000 | ~$803 (1 GPU) | $820-$1,290 | $700-$3,100 |
| 50,000 | ~$1,731 (2 GPUs) | $4,100-$6,450 | $3,500-$15,500 |
| 100,000 | ~$2,744 (3 GPUs) | $8,200-$12,900 | $7,000-$31,000 |
| 300,000 | ~$7,972 (9 GPUs) | $24,600-$38,700 | $21,000-$93,000 |
Self-hosted figures use on-demand GPU prices plus one CPU instance and Twilio inbound minutes. Commitments cut the GPU line by more than half. Managed figures exclude concurrency fees: at 300,000 minutes, peak is roughly 43 simultaneous calls, which adds about $360/month on Vapi's Core plan ($29 plus 33 extra lines × $10) and about $184 on Retell (23 extra lines × $8).
The Retell range is deliberately wide because it is the honest one: a GPT-5 nano stack at $0.0016/min LLM cost produces a very different bill than a GPT-5.5 fast-tier stack at $0.32/min.
Where the Break-Even Sits
On infrastructure alone, one self-hosted stack costs about $718/month plus $0.0085/min for telephony. Against Vapi at $0.082/min, it pays for itself at roughly 9,800 minutes/month. Against Retell's $0.11/min example stack, at roughly 7,100 minutes/month. Against Retell's $0.07 floor, the break-even stretches to about 11,700 minutes, but that floor assumes a cheap LLM and basic voice, which may not meet your quality bar.
But infrastructure isn't the whole bill. Someone has to build the agent, keep GPU drivers and models current, and answer the page when it breaks. Put a number on that time: every $1,000/month of engineering pushes the break-even out by roughly 10,000-14,000 minutes. That's the honest reason most teams start on a managed platform and move later.
Volume decides the bill; peak concurrency decides the servers. A call center with flat traffic all day gets far more out of each GPU than one where every call lands between 9 and 11 a.m.
Costs That Don't Show Up Per Minute
Managed platforms
- Concurrency fees once you pass the included lines. Vapi includes 4 (10 on Core), Retell includes 20.
- Compliance: Vapi's HIPAA add-on is $2,000/month, and Zero Data Retention needs a paid support package. Retell's BAA is self-signed with no separate fee listed, which makes it the cheaper legal path for a clinic.
- Premium voices and frontier models can double or quadruple the per-minute rate. A GPT-5.5 fast-tier LLM at $0.32/min costs about 200 times a GPT-5 nano stack at $0.0016/min.
- Twilio add-ons if you bring your own carrier: Conversation Relay adds $0.07/min, call recording $0.0025/min, and partial minutes round up.
Self-hosted
- Engineering time to build and run the stack, which is usually the biggest line.
- Idle GPUs overnight, unless you scale down outside business hours.
- Load testing and monitoring before you trust it with real callers.
- A model quality gap if one GPU has to carry everything. Faster-whisper serializes at the kernel level, so concurrency is a real constraint, not a rounding error.
- Spot interruption risk if you chase the cheapest GPU price. A >20% interruption rate on g6.xlarge spot is not compatible with live phone calls without fallback capacity.
Which One to Pick
Vapi
Best for developers who hand-pick every provider
Choose Vapi if you're a developer who wants to hand-pick every provider and already have model API keys. It offers the lowest platform fee and the most freedom to mix models. Watch the HIPAA add-on as you grow: at $2,000/month, it costs more than two self-hosted GPU stacks before a single call.
Retell
Best for a bundled engine with generous concurrency
Choose Retell if you want a bundled engine with generous concurrency and a self-signed BAA. Transcription is included in the voice engine, 20 concurrent calls come standard, and the BAA is self-signed on pay-as-you-go, with no separate fee listed. Choose your LLM carefully: it's the biggest swing factor, and the gap between GPT-5 nano and GPT-5.5 is 100×.
Self-hosted on AWS
Best past ~10,000 min/month, or when data must stay in your account
Choose self-hosted if you're past ~10,000 minutes/month or data must stay in your account. You get a fixed bill that grows with peak concurrency, not minutes. The LLM runs on vLLM in your own VPC. Budget for the engineering time honestly: it's the line item that kills the business case if you pretend it's free.
Is Self-Hosting an AI Phone Agent Worth It?
My view: start on a managed platform, prove your agent actually answers calls well, and watch your monthly minutes. Once you're steadily past 10,000 minutes a month, or compliance add-ons make the managed bill uncomfortable, price out a self-hosted stack against your real traffic. For the GPU and LLM layer, a pre-configured vLLM inference server in your own AWS account lets you pay for the GPU, not for every minute.
Frequently Asked Questions
How much does an AI phone agent cost per minute?
It depends on the platform, model, and voice. Managed platforms range from roughly $0.07 to $0.31/min all-in. Self-hosted runs a fixed GPU bill plus telephony, which works out cheaper only at sustained volume.
How much does Vapi cost?
$0.05/min platform fee plus provider costs at pass-through, working out to $0.082-$0.129/min all-in on Vapi's own example. HIPAA is $2,000/month extra; Zero Data Retention comes with any paid support package, from $29/month.
How much does Retell AI cost?
$0.055/min voice engine plus LLM and voice, with a published all-in range of $0.07-$0.31/min. The pricing page's example stack lands at $0.11/min before telephony. A self-signed BAA is available on pay-as-you-go, with no separate fee listed.
Is it cheaper to self-host an AI phone agent?
At sustained volume (roughly 7,000-10,000 minutes/month on infrastructure alone), yes. But engineering time pushes the real break-even higher. Compliance add-ons can move it earlier.
What does a self-hosted AI phone agent cost on AWS?
About $587.50/month for one g6.xlarge GPU instance on-demand, plus ~$130/month for CPU, plus $0.0085/min for Twilio inbound. A three-year reserved GPU commitment cuts the GPU line to about $269.66/month.
Do Vapi and Retell charge for concurrent calls?
Yes. Vapi includes 4 lines on its free tier and 10 on the $29/month Core plan, and charges $10/month per extra line. Retell includes 20 and charges $8/month per extra line.
Is the quality the same when you self-host?
Open-weights models have improved dramatically, and Kokoro is a capable voice. But frontier models and ElevenLabs voices still lead on complex tasks and naturalness. Test before committing, and remember that faster-whisper concurrency is limited by GPU kernel serialization, not thread count.
Run the Agent's LLM on Your Own GPU
Launch a pre-configured vLLM inference server in your own AWS account and pay for the GPU, not for every minute.
Deploy vLLM on AWS Marketplace