Retell AI pricing in 2026, the real cost per minute with add-ons included

Retell AI costs $0.07 to $0.31 a minute on its pay as you go plan, with no monthly fee, $10 of free credits and 20 concurrent calls included, according to its pricing page on October 18, 2026. That range is the sum of four lines: $0.055 a minute of voice infrastructure, a voice from $0.015 to $0.10, a model billed per minute from $0.0016 to $0.64, and $0.015 a minute of telephony when the call goes through a Retell number. A typical web agent on a recommended model and an ElevenLabs voice comes to about $0.17 a minute once two common add-ons are counted.

This page rebuilds the bill line by line, works out one minute of conversation, then the monthly cost at 1,000, 10,000 and 100,000 minutes. It ends by pricing the same conversation in your own back end, where the provider fees usually come to a fraction of the Retell bill, and by naming the cases where Retell is still the better deal.

Retell AI pricing in short

• Retell AI charges per minute of call, counted to the second, with no subscription on the pay as you go plan.
• The model is a per-minute line like the others, so a minute of silence costs the same as a minute of fast talk.
• Concurrent calls past 20, phone numbers and knowledge bases past the tenth are monthly fees on top.

How the per-minute rate is built

Retell bills each component of a call as its own line, then adds them. Every price below comes from retellai.com/pricing as read on October 18, 2026.

Voice infrastructure, $0.055 a minute

Retell charges $0.055 a minute on every call for what it calls its voice engine: the orchestration, turn-taking and interruptions. The pricing page has no separate line for speech to text, so transcription is part of this rate. The rate stays the same whichever model and voice you pick.

The voice, $0.015 to $0.10 a minute

Retell’s own voices and the voices from Cartesia, OpenAI, Inworld, MiniMax and Fish all cost $0.015 a minute. ElevenLabs voices cost more and depend on the model:

Retell bills the voice for every minute of the call, whoever is speaking. ElevenLabs sells Flash directly at about $0.04 for each minute the agent speaks, and charges nothing while the user speaks or the line is silent.

The model, $0.0016 to $0.64 a minute

Retell lists more than thirty models from OpenAI, Anthropic and Google, each with a price per minute. Its four recommended models are GPT-5.6 Terra and Claude 5 Sonnet at $0.064, GPT-5.4 at $0.080 and GPT-5.5 at $0.16. The cheapest is GPT-5 nano at $0.0016 and the most expensive standard price is GPT-6 Astra at $0.32. Most OpenAI models also have a fast tier that doubles the price, which takes GPT-6 Astra to $0.64 a minute.

Speech to speech models replace the model and the voice with a single line: GPT Realtime mini and GPT Realtime 2.1 mini at $0.07, GPT Realtime and GPT Realtime 1.5 at $0.345, GPT Realtime 2 and 2.1 at $0.38.

Retell’s pricing estimator also offers a “Custom LLM” option, for a model you host behind a WebSocket that Retell connects to. The pricing page gives it no price. Third-party guides such as eesel’s breakdown describe the model line as free in that case, so you would pay the infrastructure, the voice and your own model provider. Ask Retell to confirm it before you budget on it.

Telephony, $0.015 a minute

A call through a number bought from Retell costs $0.015 a minute in every country Retell lists, across North America, Europe and Asia. Telephony is free on your own SIP trunk and on a web call from the browser SDK. Retell bills only the telephony for the part of a call transferred to a human.

Why the advertised range stops at $0.31

The low end, $0.07, is the infrastructure plus a $0.015 voice, with a nearly free model and no telephony. The high end, $0.31, is $0.055 of infrastructure, an ElevenLabs v3 or v4 voice at $0.10 and GPT-5.5 at $0.16. The price list goes further than that range: GPT-6 Astra on the fast tier with an ElevenLabs v4 voice and a Retell number comes to $0.81 a minute, before any add-on.

Add-ons and fixed lines

Per-minute add-ons stack on top of the rate:

Other lines are billed by the month:

• 20 concurrent calls are included, and each extra one costs $8 a month, added or removed at any time.
• A Retell phone number costs $2 a month, and a verified number $10 (the page calls it monthly in one section and a one-time fee in another).
• SMS on a Retell number costs $20 a month.
• Knowledge bases cost $8 a month each past the first ten.
• Enterprise is a custom contract with concurrency from 50 calls, SSO and 24/7 support. A dedicated server and dedicated implementation support come as paid options.

Retell counts each call to the second and adds the seconds up over the month before billing. Failed connections are free. Retell bills silence and hold time, and bills voicemail only while the agent is on the line.

A minute of conversation, then the monthly bill

Take a voice agent inside a web app, answering from a product knowledge base. It runs on GPT-5.6 Terra, one of Retell’s recommended models, speaks with ElevenLabs Flash, and uses advanced denoising because people talk to it from open-plan offices. The call starts from the browser, so there is no telephony.

For the monthly bill, multiply that rate by the minutes, then add the extra concurrency. Concurrency depends on when people call. In this example, calls arrive over 22 working days of 8 hours, with three times the average load in the busiest hour. At 100,000 minutes that means about 29 calls at once, 9 above the free 20, so $72 a month.

The same agent answering phone calls on one Retell number costs $186, $1,842 and $18,474 a month at those volumes. Retell’s estimator on its pricing page runs the same multiplication with its own defaults, which come to $0.11 a minute before add-ons.

Why Retell bills the model per minute, and what that changes

Model providers bill tokens. Retell resells those tokens as a flat price per minute of call, set per model. For GPT-5.6 Terra, that is $0.064 a minute, while OpenAI sells the model at $2 per million input tokens and $12 per million output tokens.

The agent sends the system prompt and the whole history to the model again on every turn, so the token cost of a minute depends on how many turns fit in it and how long the prompt and history are. With a 1,500 token prompt and about 50 tokens of conversation per turn, a ten minute call of 40 turns sends about 100,000 input tokens, so 10,000 a minute. Here is what a minute costs on GPT-5.6 Terra at OpenAI’s list price, against Retell’s $0.064:

On most calls, Retell’s flat rate is above the token cost. A user who thinks before answering, reads something on screen or puts the call on hold costs the full minute on Retell, while the same minute costs almost nothing in tokens. A fast back-and-forth with a long prompt and many tools can cost more in tokens than Retell charges. Prompt caching lowers the token cost further, since every request starts with the same prompt.

Retell bills the voice the same way. Its $0.040 for ElevenLabs Flash covers every minute of the call, while ElevenLabs bills only the characters the agent speaks. When the agent talks half the time, buying the voice directly costs about $0.020 per minute of call, half of Retell’s price.

Running the pipeline in your own back end

The alternative to a per-minute platform is to call speech to text, the model and text to speech from your own server, on your own keys. Each provider then bills you for what it processed: audio seconds for transcription, tokens for the model, characters for the voice.

Here is the same minute with the same model and the same voice, plus OpenAI’s gpt-4o-transcribe for transcription. The user speaks for 25 seconds, the agent for 30, and the pipeline sends audio to transcription only while the user speaks.

At 1,000, 10,000 and 100,000 minutes, the provider fees come to about $44, $440 and $4,400 a month. Those figures leave out the server, the knowledge base and denoising, and the engineering time, which are now yours. A cheaper model or a cheaper voice lowers the provider fees further. Your choice of providers changes the total more than any orchestration decision.

On your own server, you run what Retell ran for you: the hosting, turn-taking and interruptions, call history, transcripts and a dashboard for whoever reviews calls, plus phone numbers if you need them. Micdrop is an open source TypeScript library for real-time voice conversations. It takes care of turn-taking and interruptions. @micdrop/client handles the microphone, voice activity detection and playback in the browser, while @micdrop/server runs the conversation in your Node server with the providers you choose. Micdrop is MIT licensed, so the library itself is free. This server runs the same model, voice and transcription as the worked example:

import { MicdropServer } from '@micdrop/server'
import { OpenaiAgent, OpenaiSTT } from '@micdrop/openai'
import { ElevenLabsTTS } from '@micdrop/elevenlabs'
import { WebSocketServer } from 'ws'

const wss = new WebSocketServer({ port: 8081 })

wss.on('connection', (socket) => {
  new MicdropServer(socket, {
    firstMessage: 'Hi, what can I help you with?',
    agent: new OpenaiAgent({
      apiKey: process.env.OPENAI_API_KEY,
      model: 'gpt-5.6-terra',
      systemPrompt: 'You answer questions about our product.',
    }),
    stt: new OpenaiSTT({
      apiKey: process.env.OPENAI_API_KEY,
      model: 'gpt-4o-transcribe',
    }),
    tts: new ElevenLabsTTS({
      apiKey: process.env.ELEVENLABS_API_KEY,
      voiceId: process.env.ELEVENLABS_VOICE_ID,
      modelId: 'eleven_flash_v2_5',
    }),
  })
})

You can run a first call in about five minutes. Wiring Micdrop into an existing application takes hours rather than days. Transcript storage, the call history, the dashboard, telephony and hosting take longer, since you build them yourself. Micdrop has a single maintainer and no support contract. Products like Raconte.ai, built by Micdrop’s maintainer, and Cibli run on it in production. If you want to switch, you can compare Retell AI and Micdrop on architecture and migration.

When Retell AI is worth the price

Retell is the better deal when the product is phone calls. Retell already sells phone numbers across North America, Europe and Asia, with calls at $0.015 a minute, and has built SIP trunking, warm transfer, IVR navigation, batch dialing and branded caller ID. Building the same telephony yourself would cost more than Retell’s per-minute margin for a long time.

Retell is also the better deal when you need compliance certifications this quarter. Retell holds SOC 2 Type 1 and Type 2, signs a GDPR data processing addendum and lets you sign a HIPAA business associate agreement online. Its compliance documentation states that Retell runs all its services outside the European Union, which matters if your data has to stay there.

Retell also makes sense at low volume. At 1,000 minutes a month, the gap with your own pipeline is about $125, less than a day of engineering. At 100,000 minutes it is about $12,600 a month, enough to pay for the hosting, the dashboard and the person who maintains them. The 20 included concurrent calls are enough for most products until well past 10,000 minutes a month, so extra concurrency only adds to the bill at high volume.

Bland also bills per minute, at $0.14 with the model included, but runs only its own models. Moving a web agent off Bland brings back the choice of providers. Synthflow has moved to annual contracts from $30,000. You can also compare every voice agent platform on price and hosting.

Frequently asked questions

Retell AI’s price is easy to read once you split it into lines: $0.055 of infrastructure, a voice, a model and telephony, then add-ons and concurrency. For a phone operation or a compliance deadline, you pay for telephony and certifications Retell already has in place. For a voice feature inside a web app at volume, the same providers called from your own server cost a fraction of the Retell bill, once you have built the hosting and the dashboard yourself. Micdrop runs the conversation inside your Node server. You can start a first call in five minutes.