Vapi vs Retell AI: The Real Per-Minute Cost

Vapi advertises $0.05 a minute and Retell AI advertises $0.07 to $0.31, but those two numbers are not the same kind of number. The Vapi figure is a platform fee with the language model, speech and telephony billed at cost on top. The Retell figure already contains them. Put both on the same component stack and a comparable minute is about $0.24 on Vapi and $0.245 on Retell. The real difference is not the rate. It is concurrency pricing, whose API keys you use, and how much of the agent you want to own.

Every comparison of these two starts by putting $0.05 next to $0.07 and declaring a winner. That is the one thing you cannot do with these pricing pages, because they are describing different products: Vapi sells orchestration and hands you the bill for everything else, Retell sells the whole call. Below is what each page printed on August 20, 2026, the same workload priced on both, and the four things that actually decide it.

Is Vapi cheaper than Retell AI?

Only on the advertised number. Vapi charges a $0.05 per minute platform fee on its Build plan and bills speech to text, the language model and text to speech at cost, or $0 if you bring your own API key. Retell AI charges $0.07 to $0.31 per minute with those components included. Add the missing pieces to Vapi and the two finish within half a cent of each other.

Both figures were read off each vendor own pricing page on August 20, 2026. We keep the wider market view, including Bland AI, ElevenLabs and Synthflow, on our AI voice agent pricing page.

What does Vapi actually cost per minute?

Vapi publishes two plans. Build is usage based at $0.05 a minute for calls and $0.005 a message for SMS and chat. Scale is an annual contract at volume based pricing, with no figure shown. Concurrency is 10 lines included, then $10 per line per month. Call history on Build is retained for 14 days, and chat history for 30.

The line that decides your bill is the one that says the model providers are billed "at cost", with $0 owed to Vapi if you bring your own API key. Vapi publishes no price for speech to text, the language model or text to speech, which means the pricing page genuinely cannot tell you what a call costs. That is not evasive, it is the design: you are buying orchestration and paying your own inference bill.

For a team that already buys inference at volume, this is the cheapest floor in the category. You keep your negotiated model rates and pay Vapi $0.05 a minute plus $10 a line to route calls. For a team that does not, the missing components are the majority of the cost and the $0.05 tells you very little.

What does Retell AI actually cost per minute?

Retell publishes $0.07 to $0.31 a minute and, unusually, prints the stack that produces the range: $0.055 for Retell voice infrastructure, $0.16 for the language model on its GPT 5.5 example, $0.015 for text to speech, and $0.015 for US Twilio telephony. Those four add to $0.245.

Read that again, because it is the single most useful number in this comparison and Retell is the only vendor that gives it to you. The spread from $0.07 to $0.31 is almost entirely which model you point the agent at. The infrastructure, speech and telephony together are $0.085 a minute and barely move. The model is everything from a couple of cents to a quarter.

Retell also meters a few things separately: 20 concurrent calls are free and additional concurrency is $8.00 per month each, branded calling adds $0.10 per outbound call, a verified phone number is $10 per number per month, and knowledge bases are free for the first 10 then $8 each per month. On a US outbound program that rotates numbers to protect answer rates, those number fees are recurring rather than incidental.

The same 1,000 minutes, priced on both

Here is one workload run through both pricing pages: 1,000 connected minutes a month, 20 concurrent lines, US telephony. Where Vapi passes a component through at cost we use the Retell published price for that component, because it is the only published US figure available. That substitution is our arithmetic, not a Vapi rate.

Two things fall out of that. At modest volume Retell is cheaper, because 20 free concurrent calls is worth $100 a month against Vapi line pricing and swamps the half cent difference in the per minute stack. And Vapi only pulls ahead when you bring your own keys, which moves $160 off this invoice and onto a model provider invoice you still have to pay. It is a genuine saving only if your negotiated rate is below the reference, which is exactly the case for teams already running inference at scale.

Scale the same comparison to 5,000 minutes and 50 lines and the concurrency gap widens: Vapi wants 40 extra lines at $10, or $400 a month, while Retell wants 30 extra at $8, or $240. Concurrency, not the per minute rate, is the line that moves most as an outbound program grows.

Concurrency is the spec an outbound team hits first

Per minute pricing gets all the attention because it is on the front of the page, but a cold calling program is capacity constrained long before it is budget constrained. If your connect rate sits near the 5.4 percent average that Gong reported across roughly 300 million calls, most of your dialing time is ring tone, and the way you buy back that time is more simultaneous lines.

That makes the concurrency line a capacity purchase, not an add-on. Vapi at $10 a line and Retell at $8 a concurrency are both cheap next to a human rep, but they are the number to model at your target volume rather than your pilot volume. It is the same structural point we make about multi line dialing in parallel dialer vs predictive dialer: the line count is the capacity spec, and the compliance obligations scale with it.

What you still have to build on either platform

Both of these are developer platforms. The invoice is the smallest part of what they cost you.

• Prompt and flow engineering. Objection handling, qualification logic and booking flows are yours to write, test and maintain against live calls.
• Telephony and number reputation. Numbers get flagged. Somebody has to monitor answer rates, rotate numbers and remediate spam labels, which we covered in how to avoid spam likely on outbound calls.
• Compliance plumbing. DNC scrubbing, calling windows, consent records and retention are not features on either pricing page. Under 16 CFR 310.5(a) the per call record you keep for five years includes the script used, the caller ID you transmitted and the disposition of the call, which a 14 day call history does not satisfy on its own. The calling window itself is fixed at 8 a.m. to 9 p.m. local time at the called person location under 16 CFR 310.4(c). More on that in is AI cold calling legal.
• Security. An agent that can look up records, transfer calls and write to your CRM is an agent with tools, and a caller can try to talk it into misusing them. If you are building rather than buying, the prompt injection and tool permission layer is part of the job too.
• CRM write back. Getting the transcript, disposition and outcome into your CRM reliably is integration work on both platforms.

None of that shows up in a per minute comparison, and it is usually a larger number than the platform bill. We put rough figures on it in AI vs human SDR cost.

Which one should you pick?

Pick Vapi if you have engineers, you already buy inference at volume, and you want maximum control over the model, the voice and the call flow. Bringing your own API keys is the feature you are actually paying for, and it is the only route in this comparison to a genuinely lower floor.

Pick Retell AI if you want a predictable bill and a shorter build. Twenty free concurrent calls, one published stack, and no separate model account to manage make it the easier of the two to forecast, and its transparency about component costs is unmatched by anyone else in the category.

Pick neither if outbound calling is the job rather than the project. Both platforms sell you the ability to make a call; neither sells you booked meetings. If nobody on your team wants to own prompt engineering, number reputation and a five year records program, a managed AI SDR prices the whole outcome instead. That is the comparison we lay out on best AI cold calling software, and the two alternative pages worth reading next are Vapi alternative and Retell AI alternative.

Frequently asked questions

Is Vapi cheaper than Retell AI? Not once the components are added. Vapi $0.05 is a platform fee with the model, speech and telephony billed at cost, while Retell $0.07 to $0.31 includes them. On the same stack both land near $0.24 a minute, and Retell 20 free concurrent calls make it cheaper at low volume. Vapi wins only if you bring your own API keys.

What is the cheapest way to run an AI voice agent? Vapi on Build with your own API keys, at $0.05 a minute plus $10 per concurrent line above 10, is the lowest platform floor available in August 2026. You then pay your own model provider directly, so the saving is real only if your negotiated inference rate beats what the platforms pay.

Why does Retell AI publish a range instead of a price? Because the language model dominates the cost and you choose it. Retell own breakdown puts voice infrastructure at $0.055, text to speech at $0.015 and telephony at $0.015, which is $0.085 combined, while the model on its GPT 5.5 example is $0.16. Swapping models is what moves you between $0.07 and $0.31.

Do Vapi or Retell AI handle TCPA and DNC compliance? No. Neither pricing page lists DNC scrubbing, consent capture, calling window enforcement or the five year record retention that 16 CFR 310.5(a) requires. Those are yours to build and operate, which is a real cost to add before comparing per minute rates.

Pricing on both platforms moves quickly. Every figure here was read on August 20, 2026, and the full market table, updated on the same date, sits on AI voice agent pricing.