Retell vs Vapi: Which AI Phone Agent Platform Fits Your Team?

Retell AI and Vapi both host AI phone agents and sell conversational AI for the phone first. You pick the models, connect a phone number, and the platform runs STT, the LLM, and the voice for every call, plus webhooks that tell your systems what happened.

The Retell vs Vapi choice is mostly about who builds the agent. Retell is an AI call center with a visual agent builder for operations teams. Vapi is the developer platform for voice AI agents and expects your engineers to configure it in code.

All feature and pricing information in this comparison comes from each vendor's official website, docs, and pricing pages, as of October 26, 2026.

What is Retell AI?

Retell AI runs the telephony, speech, and language model layers so you configure an agent instead of wiring providers together. Each agent uses a single prompt or a conversation flow, with a knowledge base and integrations built in.

You can build from the dashboard with no code or go fully programmatic with the API, Node.js and Python SDKs, and an MCP server. A drag-and-drop canvas lays out each step of the call, and an AI copilot called Conductor builds, tests, and refines the agent with you.

Agents go live for inbound calls, outbound campaigns, web, or SMS, and supervisors can monitor calls live, listen in, and take over the moment a caller needs a person. Operations teams that want an AI call center without a full engineering build are the typical users.

What is Vapi?

Vapi AI is a developer platform for voice AI agents. At its core it is an orchestration layer over three modules, the transcriber, the model, and the voice, and you choose each one from dozens of providers, including OpenAI, Anthropic, Deepgram, and ElevenLabs.

The platform handles the infrastructure while your team keeps assistants, squads, tools, and tests as files in git. Squads orchestrate several assistants that hand a call to each other with context preserved. Vapi retired its visual Workflows builder in August 2026, so the node-graph option no longer runs.

The dashboard still gets a first agent live in about 5 minutes, and the platform can be white-labeled inside your own product. Engineering teams that treat the agent like software are the typical users.

Side by side at a glance

The table lines up both platforms row by row, and the sections below break each row down in detail.

Key terms in this comparison

The comparison below leans on these terms, so knowing what each one means makes it easier to weigh Retell and Vapi feature by feature.

• STT, TTS, and the LLM: the three layers of a phone agent. Speech-to-text turns the caller's audio into text, the large language model decides what to say, and text-to-speech turns the reply into a voice.
• Latency: the pause between the caller finishing a sentence and the agent starting its reply. Under about a second, the call feels like a normal conversation.
• Turn-taking and endpointing: how the agent knows the caller has finished talking and it is time to reply. Endpointing is the detection step that makes that judgment.
• SIP trunking and phone numbers: SIP trunking connects a platform to your existing carrier or phone system over the internet, so you can bring the numbers you already own instead of renting new ones.
• Warm transfer and IVR: a warm transfer hands a caller to a person with context instead of a blind handoff. An IVR is the press-1-for-sales phone menu an outbound agent may need to navigate.
• Concurrency: how many calls your account can run at the same time. The limit matters most during your busiest hour, not your average one.
• Webhooks and Server URL: two ways a platform talks to your systems. Webhooks push call events to your code as they happen, and a Server URL is the endpoint your code exposes so the platform can ask it to run tools.

Building agents: agent builder vs code

What the numbers mean

• Who builds: Retell suits a team where an operations lead designs the conversational AI and a developer wires up the systems. The canvas lays out each step of the call, including functions, transfers, DTMF, and SMS, and Conductor drafts and tests changes with you. Vapi suits a team that treats the agent like software, with a git workflow that puts every assistant through code review.
• Integrations: Retell syncs contacts and call data with Salesforce, HubSpot, Dynamics 365, GoHighLevel, and Zoho, while Calendly and Cal.com handle appointment booking, and webhooks push real-time call events so a CRM updates without polling. Each integration tool call times out after 3 to 14 seconds, which only matters if your back-end systems respond slowly.
• Tool calling: on Vapi, tool calling for advanced features runs through a Server URL, a webhook endpoint your own code answers, and a visual Tools section lets you reuse the same tools across assistants. You write more plumbing, and you keep full control of it.
• Multi-step logic and testing: with Workflows gone, multi-step logic on Vapi lives in Squads of assistants that hand off to each other. Simulations run an AI tester against each assistant to catch broken flows, bad handoffs, and voice issues before real callers reach them.

Verdict

• Pick Retell if an operations lead will own the agent day to day, because the canvas and Conductor could turn every script change into a dashboard edit instead of an engineering ticket.
• Pick Vapi if your engineers want the agent in version control, because config files in git, reusable tools, and Simulations may fit a team that already reviews every change as code.

Latency, turn-taking, and voice quality

What the numbers mean

• Latency: Retell quotes about 600ms, roughly the pause a person leaves before replying. Vapi publishes several latency figures across its site and docs, from under 500ms on average to a typical 800ms end to end. Each vendor measures its own number, so treat the published figures as ballparks and time both platforms on your own calls. On both, the model you choose moves latency more than the platform does.
• Turn-taking: Retell runs its own turn-taking model that knows when to stop and when to listen, and its agents handle mid-sentence interruptions instead of waiting for a full pause. On Vapi, interruption handling is configurable: you set how many words a caller must say before the assistant stops talking, and for English its docs recommend LiveKit's smart endpointing to judge when a caller has finished.
• Voice quality: Retell offers its own voices built from real performance data, alongside Cartesia, ElevenLabs, OpenAI, and other TTS providers, and for speech-to-text it supports Deepgram, Azure, Soniox, and AssemblyAI. On Vapi, voice quality depends on the TTS provider you pick, and its own curated voices span 40+ languages with automatic language detection.

Verdict

• Pick Retell if you want natural turn-taking without tuning, because its own interruption model could get a front-desk or support line sounding conversational on day one.
• Pick Vapi if your team wants to tune the conversation itself, because the word threshold and smart endpointing settings put those dials in your engineers' hands.

Telephony, transfers, and outbound calls

What the numbers mean

• Carriers and numbers: Retell connects to Twilio, Vonage, or any phone system through SIP trunking. Vapi also works with any SIP provider and lists Twilio, Vonage, and Telnyx rates that the carrier bills directly, so on Vapi the telephony line lands on the carrier's invoice rather than the platform's.
• Transfers: a Retell warm transfer keeps the agent on the line until a person picks up, and it can whisper context to that person or make a three-way introduction. An agentic version screens the destination before it bridges the caller. Vapi's warm transfer that speaks a call summary works on Twilio only, and its most flexible mode carries an Experimental label, which matters if a reliable handoff to a human is a launch requirement.
• Outbound calling: Retell builds in outbound dialing with batch calling from a CSV, and its agents can detect an IVR menu and press digits to reach the right extension during a live call. Retell has no per-call delay setting, so concurrency sets the pace of a campaign. Vapi's calling campaigns also run from a CSV and can be scheduled up to seven days ahead, but you can't save a campaign as a draft.

Verdict

• Pick Retell if callers often need to reach a person, because the warm transfer with whispered context and the AI screening option could make those handoffs smoother.
• Pick Vapi if you already run numbers on your own SIP carrier and plan outbound work ahead, because any SIP provider connects and campaigns can be scheduled days in advance.

Pricing and concurrency

What the numbers mean

• Rate structure: on Vapi, pricing starts with a platform fee for hosting, and every model passes through at cost with no markup, so the bill tracks whatever stack you pick. Retell's advertised rate has no separate platform fee, but a $0.055/min Voice Infra charge sits inside every minute, next to the LLM, TTS, and telephony costs, and its published range runs up to $0.31 per minute depending on the models and add-ons you choose, with no annual contract.
• How billing runs: Retell bills to the nearest second, the agent fee stops once a call transfers, and silence still counts because speech-to-text keeps listening, so a quiet caller costs the same as a talkative one. Add-ons such as PII redaction carry their own per-minute price.
• Capacity and support: concurrency is where the bills diverge: a busy inbound line hits Vapi's extra-line pricing sooner than Retell's. Vapi's base plan also includes 14 days of call data, and paid support packages start with Core. An uptime SLA starts with the Pro package at 99%, and Premier raises it to 99.9%.
• Budgeting: for a rough budget, multiply your expected monthly minutes by each platform's per-minute total from its pricing calculator, then add the extra capacity your busiest hour needs.

Verdict

• Pick Retell if you want one bundled rate and headroom for busy hours, because 20 included lines could carry a mid-size inbound queue before any capacity fees start.
• Pick Vapi if you want to see and control every model cost, because at-cost pass-through may work out cheaper once your engineers tune the stack, especially at low call volumes.

Compliance and data handling

What the numbers mean

• Certifications: Retell's docs list HIPAA, GDPR, and SOC 2 Type 1 and Type 2. Its homepage lists ISO 27001, which its compliance docs do not. Vapi lists SOC 2, HIPAA, PCI, and GDPR. These acronyms mostly matter to your security reviewer, and the two lists overlap on HIPAA, GDPR, and SOC 2.
• HIPAA in practice: Retell signs a HIPAA agreement with no additional fee. On Vapi, HIPAA is a $2,000 a month add-on, and turning on HIPAA mode limits the organization to compatible model, voice, and transcriber providers, so a US healthcare line starts cheaper on Retell.
• Where the data lives: enterprise teams can deploy Retell on their own infrastructure, and Retell has zero data retention agreements with its LLM, speech, and voice subprocessors, which means those providers don't keep your call data. Retell doesn't currently operate services in the EU, which rules it out for EU data residency, while Vapi has no on-premise option and its EU region comes only with the Premier package. An EU-based contact center has one option between the two.

Verdict

• Pick Retell if you run a US healthcare line or calls can't leave your own servers, because the no-fee HIPAA agreement and on-prem deployment could remove two budget lines at once.
• Pick Vapi if EU data residency is a legal requirement, because its Premier EU region is the only way to keep call data in the EU between these two platforms.

Put your Retell or Vapi agent on a screen

Phone agents built on either platform often end up answering on a website or in an app. That's where LemonSlice fits: a real-time interactive AI avatar layer on top of your voice agent. It works with any TTS and any LLM, so the prompt, tools, and voice you tuned for calls stay the same.

LemonSlice works best for customer support, healthcare intake, sales training, tutoring, and lead qualification. Whatever your agent handles, here's what comes with it:

• Any presenter from one photo: a brand mascot or a photo of your support lead goes live with no training or video shoot.
• Response speed: LemonSlice adds minimal latency on top of your existing voice agent. Its global fleet of GPUs runs calls from the nearest region.
• Let your coding agent add it: install the LemonSlice skill, then ask Cursor, Claude, or Copilot to add a video avatar to your voice app. The skill scaffolds a starter app with keys, the voice cascade, an image, and a ringing UI.
• Wire any stack: LemonSlice's WebSocket endpoints drive an avatar from any audio stack, which suits an agent whose audio you already stream, and you keep your own speech-to-text, LLM, and TTS. You can walk through the full flow in our docs.

The WebSocket connection is the whole integration:

const socket = new WebSocket(websocketAddress);

{
  "command": "audio",
  "audio": "<base64_pcm16_mono>",
  "sampleRate": 16000,
  "encoding": "PCM16"
}

Code sample from lemonslice.com/docs/websocket

On Enterprise plans, sessions stay live for up to 24 hours, long enough for an interactive museum exhibit. Developers can start with as little as $8 a month, and every plan includes unlimited avatars and API access. Add your first interactive AI avatar to your Retell or Vapi agent.

Retell vs Vapi: which should you choose?

Match the platform to the team that will own the agent.

Choose Retell if an operations team will build and run the AI voice agent, because the drag-and-drop canvas, built-in outbound campaigns, and simulation testing don't need an engineer for every change.

Choose Vapi if your developers want to pick every provider and ship agents like code, because it passes model costs through at cost and keeps the whole configuration in git.

Still torn? Write one script for your most common call, such as lead qualification, run it on both platforms with the same LLM and voice, and compare latency, transfer behavior, and the per-minute bill.

Whichever way you go, you can give your voice agent a face with LemonSlice. It works on top of whichever platform you pick: you can create the character from a single photo, and a real-time interactive avatar answers on your website while the phone line keeps running. Responses come back with ultra low latency from LemonSlice's global fleet of GPUs, and developers can start prototyping for as low as $8 a month with API access included. Swap the platform later and the face carries over. Build your first interactive AI avatar with LemonSlice.