Retell AI Review: What It Actually Costs and Who It Suits

If you are evaluating platforms to build an AI voice agent, Retell AI will come up early in your research. It should. It is one of the more capable platforms available right now for deploying conversational voice agents on real phone lines, and unlike some competitors it is designed with developers and integrators in mind rather than just enterprise procurement teams.

Most reviews of it stop at "latency is good, voices are nice". That is true and not very useful, because the question that decides whether you deploy is what a minute of conversation costs you, and Retell does not answer that with a single number. It bills in four separate layers that you assemble yourself. This review works through those layers using the figures Retell publishes, then covers what the platform is good at, where it frustrates, and who should not use it.

I have built multiple production voice agents on Retell across real estate, home services and medical appointment booking, plus a multilingual booking agent for a European beauty salon that handles Dutch, French and English callers on the same line. The opinions here come from that work. The numbers come from the vendor.

What Retell AI Actually Is

Retell AI is voice agent infrastructure. It handles the hard parts of voice AI: low-latency speech-to-text, LLM processing, text-to-speech, and phone call management over SIP or PSTN. You bring the conversation logic, the prompts, the flow and the integrations. Retell handles the voice layer.

That distinction matters. Retell is not a no-code chatbot builder with a voice feature bolted on. It is built specifically for voice, with real-time audio processing at its core. The result is a far more natural conversation than you get from general-purpose automation tools that added voice as an afterthought.

You can connect Retell to your own LLM or use its built-in agent builder. Retell AI Custom LLM Integration covers the first route. Phone numbers are provisioned through Twilio, Telnyx or your own SIP trunk; Twilio's Programmable Voice documentation covers the telephony side and Twilio Voice Setup for AI Agents covers that step in detail. Inbound and outbound calling are both supported.

The Pricing, Layer by Layer

Retell's pricing page advertises a band of $0.07 to $0.31 per minute for voice agents, pay as you go, with $10 in free credits and no contract. That band is wide for a reason: it is the range you land in once you have chosen a model and a voice, not a rate anyone charges you directly. Four components stack up.

The LLM layer is where the range comes from. A selection from the published rate card:

Speech-to-speech models are priced separately and are considerably dearer: GPT Realtime and GPT Realtime 1.5 are $0.345 per minute, with GPT Realtime mini at $0.07.

Retell's own worked example on the pricing page assembles a call at $0.11 per minute: $0.04 LLM, $0.055 voice infra, $0.015 TTS, and $0.00 telephony because that example brings its own SIP. That is a realistic floor for a competent production agent. Swap in ElevenLabs and a Twilio number and the same call is roughly $0.15. Swap in GPT 5.5 on the fast tier and a premium voice and you are north of $0.42, well above the advertised ceiling, because the ceiling assumes you did not do that.

Use the AI Voice Agent Cost Simulator to model what your own call volume looks like once you have picked a stack.

The add-ons nobody budgets for

Individually small, collectively not:

• Knowledge base: +$0.005 / min
• Advanced denoising: +$0.005 / min
• Safety guardrails: +$0.005 / min
• PII removal: +$0.01 / min
• Batch calling: +$0.005 per dial
• Branded calling: +$0.10 per outbound call
• AI quality assurance: $0.10 / min, first 100 minutes free
• SMS: $0.01 per message

Turn on knowledge base, denoising, guardrails and PII removal on a healthcare agent and you have added $0.025 a minute before a word is spoken. Branded calling at ten cents a dial is the one that surprises outbound teams: on 2,000 dials a month that is $200 on its own, more than the conversation cost for a campaign of short calls.

The monthly line items

Pay as you go does not mean zero fixed cost. Phone numbers are $2.00 a month each. Retell SMS is $20.00 a month. Verified phone numbers are $10.00 a month. Concurrency is free to 20 simultaneous calls, then $8.00 per concurrent call per month beyond that, and knowledge bases are free to ten, then $8.00 a month each. A shop running 40 concurrent lines is paying $160 a month before any minutes.

Billing mechanics worth knowing

Retell's billing FAQ answers three questions that catch people out, and the answers are not all in your favour:

• Usage is tracked to the nearest second, not rounded up per call. Good news, and unusual.
• You are billed during silence and hold, because speech-to-text keeps running. An agent that parks a caller while it queries a calendar is still costing you money.
• After a transfer to a human, the AI fee stops but telephony continues. Long handovers are cheaper than they look but not free.

Failed calls are not billed. Voicemails are billed only for the seconds the agent was actually active. Enterprise pricing is custom and starts at 50+ concurrency, with dedicated servers, custom SSO, role-based access control and 24/7 support. Note that HIPAA no longer belongs on that list. Retell's compliance documentation now offers the Business Associate Agreement and the Data Processing Addendum for self-signing at no additional fee on the pay-as-you-go plan, alongside per-agent data retention from one day to two years and per-agent PII controls. If you were told to look elsewhere because Retell gated compliance, that advice is out of date, and the alternatives worth considering are now a shorter list than they were.

What Works Well

Latency

This is the make-or-break metric for voice AI. If the agent pauses for two seconds to think between sentences, callers hang up. Retell has invested heavily in reducing end-to-end latency, the time between a caller finishing a sentence and the agent responding. On the builds I have shipped, response latency sits consistently in the 700ms to 1,200ms range, which reads as natural conversation speed.

A well-configured Retell agent feels noticeably more responsive than most IVR systems and comparable to a slightly slow human. Callers who do not know they are talking to an AI often do not work it out.

Voice quality

Retell supports several TTS providers: its own platform voices plus Minimax, Fish, Cartesia, ElevenLabs and OpenAI. ElevenLabs through Retell gives you some of the best-sounding AI voices available, expressive and available across accents and languages, and it is the reason the TTS line jumps from $0.015 to $0.040. Whether that 2.5x is worth it depends entirely on your use case. For a dental reminder call, it is not. For a premium services brand where the voice is the brand, it plainly is.

Call handling

Retell handles the practical realities of phone calls: detecting when the caller stops speaking, managing interruptions when a caller talks over the agent, transferring to a human number, ending gracefully, and logging the transcript. None of these are trivial to build. Having them at platform level means you design conversations rather than telephony plumbing.

Webhooks and integrations

Every call fires webhooks: call started, call ended, transcript available, custom tool invoked. These integrate cleanly with Make.com, n8n or any custom backend. I use them to push call summaries and captured data into GoHighLevel, Google Calendar for booking, and Slack for real-time notifications, which is the pattern behind How to Use Retell AI. The integration surface is solid.

What Could Be Better

The rate card is the documentation

You have to read a four-layer rate card and assemble your own number before you can answer "what does this cost". That is honest pricing but it is not friendly pricing, and it means most cost comparisons you find online are quoting whichever layer the author happened to look at. Work out your own figure before you commit to anything.

The no-code builder has limits

Retell has a visual conversation flow builder that keeps improving, but it is not as capable as writing a strong system prompt and handling logic through tool calls and webhooks. If you want an agent that looks up data, makes decisions and handles edge cases gracefully, you will eventually move beyond the basic builder into a custom LLM integration or a backend server handling tool calls. How to Write a System Prompt for AI Agents is the right next read at that point.

This is not really a criticism. It is where voice AI tooling is. The no-code surface is genuinely useful for simple cases and prototyping. Production agents with real complexity need a developer or a developer-minded integrator.

Documentation gaps

The docs have improved a lot over the past year, but there are still corners where the best information lives in the Discord community rather than the official documentation. If you are building something non-standard, plan to spend time in the community or work with someone who has already hit the edge cases.

How Retell Compares

Retell is not the only option and the space moves fast. Vapi's documentation is worth reading for that platform's own account of what it does rather than taking this table's word for it. But for production voice agents that need to integrate cleanly with CRMs and booking systems, Retell is the platform I reach for most often.

Who Should Use Retell AI

Retell fits if you are a developer or working with one, need production-grade agents rather than demos, want to integrate with an existing CRM or scheduling stack, and need both inbound and outbound calling.

It fits badly if you want a completely no-code experience with no technical knowledge. The platform rewards people who understand APIs, webhooks and prompts. If that is not you, you will get further faster working with someone who already knows the platform.

For most small and mid-size businesses, the sensible move is not to learn Retell yourself but to have someone build and configure the agent, hand it over fully integrated, and maintain it. At around eleven to fifteen cents a minute all-in, a well-built agent pays for itself quickly once you weigh it against what you currently lose to missed calls or spend on a manual answering service. AI Voice Agent for Small Business walks through that ROI case properly.

The Verdict

Retell AI is good. Latency is competitive, voice quality is excellent with the right TTS provider, and the integration surface is developer-friendly without being needlessly complex. The unbundled pricing is more honest than a single blended per-minute rate, even if it makes budgeting slower.

For building AI voice agents that handle real inbound calls, qualify leads and connect to your business tools, Retell is the right foundation. The gap between a working prototype and a polished production agent is mostly conversation design and integration work, not the platform. That integration work, connecting Retell to GHL, Airtable and calendar systems for real deployments, is where most of my own build time goes.

If you want to see what this looks like for your specific business, the call flow, the integrations and the real monthly cost, book a free 30-minute call. We can scope a working agent in one conversation.

Frequently Asked Questions

What does a Retell agent actually cost per minute?

Less than the advertised band suggests at the bottom and more at the top, because the band is assembled rather than charged. The Retell layer itself is $0.055 a minute. Standard text-to-speech adds $0.015, ElevenLabs adds $0.040, telephony through Twilio or Telnyx adds $0.015 and your own SIP trunk adds nothing. The LLM ranges from $0.003 a minute for GPT 5 nano to $0.16 for GPT 5.5, or $0.345 for a speech-to-speech realtime model. Retell's own worked example lands at $0.11 a minute. A realistic production agent on a standard voice and a mid-tier model sits between eleven and fifteen cents, before add-ons.

What are the hidden costs?

The per-minute add-ons and the monthly line items. Knowledge base, advanced denoising and safety guardrails are $0.005 a minute each, PII removal is $0.01 and AI quality assurance is $0.10 after the first 100 free minutes. Branded calling is $0.10 per outbound call, which dominates the bill on high-volume short-call campaigns. On the monthly side, numbers are $2 each, SMS is $20, verified numbers are $10, and concurrency above the free 20 simultaneous calls is $8 per concurrent call per month. You are also billed during silence and hold, because speech-to-text keeps running while the caller waits.

Do I need Twilio, or can Retell provide the number?

You can buy a number inside Retell and be calling within minutes, which is the right way to prototype. Bringing your own SIP trunk is the better production answer for two reasons: it removes the $0.015 a minute telephony charge entirely, and it keeps the phone number, the call records and the carrier relationship under your control rather than inside a platform you may later want to leave. Porting a number you already advertise is far more painful than provisioning it correctly at the start.

Will callers know they are talking to an AI?

Many will not, at least not immediately, which is precisely why you should tell them. Latency and voice quality are good enough that the illusion holds for a routine booking call, but it breaks the moment the conversation goes somewhere unexpected, and a caller who feels deceived at that point is harder to recover than one who knew from the first sentence. Disclosure rules for automated callers vary by jurisdiction and have been changing, so check your own market rather than assuming.

Should I use the visual flow builder or a custom LLM?

Start with the builder and move when it stops fitting. For a booking agent with a handful of paths, the visual flow is faster to build and much easier to hand to someone non-technical. You outgrow it as soon as the agent needs to look something up mid-call, make a judgement, or recover from a caller doing something unusual. Rebuilding on a custom LLM later is not wasted work, because the conversation design you validated in the builder is the part that carries over.

Is Retell the right choice if I have no developer?

Probably not on your own. The platform rewards people comfortable with APIs, webhooks and prompts, and the gap between a working demo and something you would put on your main business line is almost entirely integration work. That does not mean avoid Retell; it means the realistic path is having someone build and hand it over configured, rather than learning telephony infrastructure as a side project.

If you would rather have this built than build it, I take on Retell AI voice agent work through Upwork and Fiverr.