Ahmad AffanAI voice
AI voice agents for business: Retell vs Vapi on cost, latency and fit (2026)
The headline price per minute is never the bill. Here is what an AI phone agent really costs, how fast it answers, and how I pick a platform.
By Ahmad AffanUpdated 3 min read
Key takeaways
- As of October 2026 Vapi lists $0.05/min for hosting and Retell starts at $0.07/min, but model, voice and telephony cost extra.
- All-in costs commonly land around $0.10 to $0.30 per minute, and the model you choose moves the price most.
- Independent real-call tests disagree on latency, from about one second to about 1.7 seconds, so measure your own agent.
- The tail matters more than the median: slow turns are the ones that make callers hang up.
- Pick the platform after you decide the call type, tools and handover rules, not before.
An AI voice agent answers the phone, listens, decides what to say or which tool to call, and speaks back, usually within a second or two. For a business it can be a receptionist that books appointments, a screener for new leads or a reminder caller. Two platforms come up in almost every conversation: Retell AI and Vapi. This guide compares them using published prices and independent tests as of October 2026, and then explains how I choose.
I have built on both sides of this: a booking platform where a phone agent talks to a custom model, Appointox, and an outbound control panel with assistants, leads and campaigns, VoxPilot.
How a voice agent works
- Speech to text converts what the caller says into text.
- A language model decides the reply or calls a tool, such as “check calendar availability”.
- Text to speech turns the reply into audio.
- Telephony carries the call to and from a real phone number.
Every step adds a little delay and a little cost, which is why platform choice affects both.
What it really costs per minute
Both platforms advertise a low platform fee. The real rate is the platform fee plus speech recognition, the language model, the voice and the phone line.
| Vapi | Retell AI | |
|---|---|---|
| Platform fee | $0.05/min hosting | From $0.07/min (voice infrastructure $0.055/min) |
| Models and voice | Passed through at provider cost | Itemised: voices from $0.015/min, LLMs from under a cent to $0.32/min |
| Telephony | Charged by the phone provider | $0.015/min on a Retell number, none over your own SIP trunk |
| Typical all-in | Roughly $0.10 to $0.42/min | Roughly $0.10 to $0.31/min |
One independent comparison estimated about $0.098/min for a mid-tier Retell setup using a small model, a platform voice and standard telephony. Another measured actual invoices on a test run and found Vapi cheaper per minute. Those figures depend on the models and voices chosen, so treat all of them as estimates and model your own call volume.
Latency: the number callers feel
People leave very little silence between turns in conversation, so anything near two seconds feels broken. Published tests give different answers depending on method:
- A caller-side benchmark of real phone calls, through August 1, 2026, measured a median of 1,558 ms for Vapi and 1,740 ms for Retell, with the slowest 5% of turns at about 2.0 s and 2.3 s.
- A smaller test of eleven calls per platform measured about 1.56 s for Vapi and 1.04 s for Retell.
- Retell’s own published metric is measured server-side and excludes the caller’s network, so it is not comparable.
How I choose between them
- Retell is quicker to configure for standard flows such as booking, with itemised pricing that is easy to forecast.
- Vapi gives a developer finer control over each component and suits custom pipelines, with more to configure.
- Custom LLM over WebSocket (supported by both) lets the agent call your own backend tools, which is what I use when booking rules are complex.
The platform is only half the project
Good voice agents are built on a call playbook, not a prompt alone. Before choosing a vendor, decide:
- One call type to automate first, for example after-hours bookings.
- The tools the agent needs: live calendar availability, pricing rules, CRM write-back.
- When it must hand over to a person, and how the summary gets to them.
- What it must never promise, such as prices or medical advice.
- What gets recorded, and how callers are informed.
If you are considering a phone agent for bookings or lead calls, see my AI voice agent service or send me the call you want to automate.
Sources
Checked on 28 October 2026. Prices and versions change, so confirm with the vendor before you commit.
Frequently asked questions
As of October 2026, platform fees start at $0.05/min on Vapi and $0.07/min on Retell, but the language model, voice and telephony are extra. All-in costs commonly land around $0.10 to $0.30 per minute depending on the models and voices you pick.
It depends on the test. One real-call benchmark found Vapi slightly faster at the median, another small test found Retell faster. Setup choices matter more than the logo, so measure your own agent.
Yes, if it is connected to a calendar or booking system through tool calls. It checks availability while the call is live and creates the booking before the caller hangs up.
Under about one second feels natural. Around two seconds feels like a dropped line, so pay attention to your slowest turns as well as the median.
