Contents
- Verdict
- What is Retell AI?
- Best for, and not for
- Strengths and limits
- Pricing
- What people actually build with Retell AI
- What Retell actually costs at small-business volume
- The gotcha: the add-ons you will actually turn on
- What breaks in production, and how you find out
- What it connects to that you already run
- Retell AI alternatives: what to pick instead, and when
- Migration and lock-in: can you get your work out?
- FAQ
Retell publishes the most detailed per-component voice pricing of anyone in this category, which makes it the easiest to forecast honestly. As of August 2026 the all-in range is $0.07 to $0.31 a minute, built from voice infrastructure at $0.055, text-to-speech at $0.015, the language model at $0.003 to $0.16 depending on which you choose, and telephony at $0.015 on US Twilio rates. Add-ons are individually priced: knowledge base, denoising and guardrails at $0.005 a minute each, PII removal at $0.01. On our 1,500-minute reference workload that is roughly $105 at the cheap end and $465 at the expensive end - and that spread is almost entirely your model choice.
What is Retell AI?
Retell AI is a platform for building conversational voice agents that handle inbound and outbound phone calls. It manages the real-time speech pipeline, combining speech-to-text, an LLM for reasoning, and text-to-speech, with a focus on natural turn-taking and low response latency. It offers a mix of a visual builder and APIs, plus telephony integration, so teams can stand up call agents for tasks like support, scheduling, and lead qualification.
Best for
Teams that want a faster path to a working phone agent with a builder plus API, without assembling the whole voice stack themselves.
Not for
Businesses that need a single predictable monthly figure. Retell's transparency is a genuine strength and the flip side is a bill assembled from a dozen line items that each move with usage.
Strengths
- The most granular published pricing in voice AI, so you can forecast to within a few dollars before building
- Model choice spans a fifty-fold price range, so cost tuning is a real lever rather than a theoretical one
- Pay-as-you-go with no minimums or contracts, plus $10 in free credits to start
- 20 concurrent calls included free, which covers peak load for most small businesses without an upgrade
- Compliance-adjacent features priced individually - PII removal at $0.01 a minute rather than a $2,000 add-on
- Ten knowledge bases included, so multi-purpose agents do not immediately incur storage fees
- AI quality assurance as an option, which is the honest way to find out how your agent is actually performing
Limitations
- The bill is assembled from many small line items, which is transparent but not simple
- The $0.07 to $0.31 range is wide enough that a careless model choice quadruples your costs
- No volume discounts published on the pay-as-you-go page, so scale does not automatically get cheaper
- AI quality assurance at $0.10 a minute nearly doubles the cost of a cheap configuration if left on
- Telephony rates vary by country and the published figure is US Twilio, so international costs need separate checking
- As with all voice agents, latency and interruption handling need real-world testing that the pricing page cannot tell you about
Retell AI pricing
Retell prices per minute, itemised by component, with pay-as-you-go billing and no minimums or contracts. Published rates as of August 2026: voice infrastructure $0.055 a minute, text-to-speech $0.015, language model $0.003 to $0.16 depending on the model selected, and telephony $0.015 at US Twilio rates, varying by country - giving an overall $0.07 to $0.31 range. Add-ons are separately priced per minute: knowledge base $0.005, advanced denoising $0.005, safety guardrails $0.005, PII removal $0.01, and AI quality assurance at $0.10 with the first 100 minutes free. Fixed monthly items include Retell phone numbers at $2, verified numbers at $10 plus a $10 one-time fee, extra concurrent calls at $8 each per month above 20 free, and knowledge base storage at $8 per base above 10 free.
| Plan | Price | What you get |
|---|---|---|
| All-in range | $0.07-$0.31/min | Pay-as-you-go, no minimums or contracts, $10 in free credits to start (as of Aug 2026) |
| Component rates | Itemised | Voice infrastructure $0.055/min, TTS $0.015/min, LLM $0.003-$0.16/min by model, telephony $0.015/min US Twilio |
| Per-minute add-ons | $0.005-$0.10 | Knowledge base, denoising and guardrails $0.005 each; PII removal $0.01; AI quality assurance $0.10 (first 100 min free) |
| Fixed monthly items | $2-$10 | Retell phone number $2/mo; verified number $10/mo plus $10 one-time; extra concurrency $8 per call/mo above 20 free; KB storage $8/mo per base above 10 free |
| Worked example | $105-$465 | 1,500 minutes a month at $0.07 to $0.31; the spread is driven almost entirely by which language model you select |
| Enterprise | Custom | Custom pricing on contact; no volume discounts published on the pay-as-you-go page |
Pricing reflects public plans as of August 6, 2026 and can change. Check Retell AI for the latest.
What people actually build with Retell AI
Retell's component pricing means the workloads that land on it are the ones where somebody sat down and worked out what a call should cost. That tends to select for high-volume, repetitive, structured calls - which is also, conveniently, where a voice agent is actually good.
Inbound appointment booking for a dental practice is the archetype. The agent answers, identifies whether the caller is an existing patient, reads the knowledge base for what the practice offers and how long each appointment takes, checks the schedule through a function call, and books. The economics work because a booking call is three to five minutes of almost entirely predictable conversation, and on a cheap model configuration that is under fifty cents of platform cost against a chair hour worth considerably more.
Outbound recall and reactivation is the same machinery pointed the other way, and it is where veterinary clinics and medical clinics get the most out of it. A nightly job pulls patients or pets overdue for a vaccine, review or hygiene visit, filters out anyone with a future booking or a do-not-contact flag, and hands a list to the agent. The agent calls in a window you choose, offers two slots, books or reschedules, and marks the record. Suppression rules matter more than the script here: a cheerful vaccine reminder for a pet that died last month does more damage than the whole campaign earned.
After-hours emergency capture for an HVAC or plumbing company uses Retell's fallback behaviour rather than its conversation design. The agent qualifies the emergency, pages the on-call technician, and - critically - has a `fallback_number` configured so that a caller who arrives when every line is busy reaches a human instead of being hung up on.
Two more that recur: a salon rebooking line that takes the client's usual service and stylist and offers the next matching slot; and an order-status line for a restaurant group where the agent reads from a knowledge base at $0.005 a minute rather than integrating with anything at all, which is often the fastest useful thing to ship.
What Retell AI is genuinely bad at
Giving you one predictable number. The itemised model that makes Retell forecastable also makes it fiddly: a dozen line items each moving with usage, four optional per-minute add-ons that a real deployment ends up wanting, and a model dropdown that swings the bill by a factor of five. If your finance function needs a fixed monthly figure on a budget line, this is the wrong shape of product and a bundled competitor is worth the premium. It is also poor, like every voice agent, at any call requiring discretion rather than information.
Bottom line: Retell AI suits high-volume structured calls - booking, recall, qualification - where the conversation is predictable enough that a cheap model handles it and the per-minute arithmetic actually matters.
What Retell actually costs at small-business volume
Retell is the only voice platform in this directory that lets you build your own per-minute rate from a published component table, so let us actually do it for our 1,500-minute reference workload. Cheap configuration, as of August 2026: voice infrastructure at $0.055 a minute, a Retell-hosted voice at $0.015, a small fast model - the table lists Claude 4.5 Haiku at $0.025 - and US telephony via Twilio at $0.015. That totals $0.11 a minute, or about $165 a month.
Expensive configuration, same call volume: voice infrastructure $0.055, ElevenLabs text-to-speech at $0.040, a flagship model - the table lists GPT 5.5 at $0.16 a minute - and the same $0.015 telephony. That is $0.27 a minute, or roughly $405 a month. Identical calls, identical minutes, two and a half times the bill. The entire difference is two dropdowns.
That is the single most useful thing to take from this page. In voice AI, model and voice selection dominate cost far more than platform choice does, and Retell is the vendor that makes this visible instead of hiding it inside a blended rate. Start on the cheap configuration, listen to real calls, and upgrade only the component that is actually letting you down - which is usually the voice, not the model.
The fixed items are small but easy to forget: a Retell phone number is $2 a month, a verified number $10 a month plus a $10 one-time fee, extra concurrent calls $8 each per month above the 20 included, and knowledge base storage $8 per base above the 10 included. Chat agents, if you use them, are billed separately at $0.001 to $0.052 per message depending on model. Signup includes $10 of free credit, which at the cheap configuration is about 90 minutes of real testing - enough to genuinely evaluate it.
| Layer | Lean config | Balanced config | Premium config | Bundled or separate |
|---|---|---|---|---|
| Voice infrastructure | $0.055/min | $0.055/min | $0.055/min | Billed by Retell, unavoidable |
| Text-to-speech | $0.015 (Retell voice) | $0.015 (Cartesia or OpenAI) | $0.040 (ElevenLabs) | Billed by Retell, priced per provider |
| Language model | $0.003 (GPT 5 nano) | $0.025 (Claude 4.5 Haiku) | $0.16 (GPT 5.5) | Billed by Retell, per minute not per token |
| Telephony, US Twilio | $0.015/min | $0.015/min | $0.015/min | Billed by Retell; $0 on your own SIP trunk |
| All-in per minute | $0.088 | $0.110 | $0.270 | - |
| Minutes at 4,000/mo | $352.00 | $440.00 | $1,080.00 | - |
| Phone number | $2.00 | $2.00 | $2.00 | Fixed monthly |
| Monthly total | $354.00 | $442.00 | $1,082.00 | - |
| Add-ons if all four are on | +$100.00 | +$100.00 | +$100.00 | KB, denoising, guardrails $0.005 each, PII removal $0.01 |
The all-in stack at 1,000 calls a month
Scale to 1,000 inbound calls averaging four minutes - 4,000 minutes - and Retell is the one platform here where every single line of the bill comes off a published page. Nothing is estimated, and nothing is left to another vendor's pricing sheet. A lean configuration lands at $0.088 a minute, a balanced one at $0.11 and a premium one at $0.27, which is $354, $442 and $1,082 a month respectively once the $2 phone number is included.
The floor is fixed and worth memorising: $0.055 voice infrastructure plus $0.015 for a Retell-hosted voice plus $0.015 US telephony is $0.085 a minute you will pay whatever else you do. Everything above that is model choice and add-ons. At 4,000 minutes the floor alone is $340, so the honest question is never whether Retell is cheap but which model you can live with.
Bottom line: At 4,000 minutes a month Retell AI runs $354 lean, $442 balanced and $1,082 premium - a three-fold spread produced entirely by the model and voice dropdowns, not by the platform.
The gotcha: the add-ons you will actually turn on
Retell's per-minute add-ons look trivial individually and are not trivial together. Knowledge base at $0.005, advanced denoising at $0.005, safety guardrails at $0.005 and PII removal at $0.01 are each a rounding error next to a $0.11 base. Turn on all four and you have added $0.025 a minute - a 23 percent increase on the cheap configuration. The catch is that these are not exotic options; they are the things a real deployment ends up wanting. Denoising because callers ring from cars, a knowledge base because the agent needs to know your opening hours, PII removal because you are handling customer details. Budget the configuration you will end up with, not the one you start with.
AI quality assurance is the outlier at $0.10 a minute after the first 100 minutes free - close to doubling the cheap configuration. It is genuinely useful for reviewing what your agent has been saying, but at that rate it is a sampling tool rather than something to leave on across all traffic. Branded Call ID at $0.10 per outbound call is similarly priced per call rather than per minute, which matters a great deal if your calls are short: on a 30-second outbound reminder, that one add-on can exceed the cost of the call itself.
There is a second dropdown most buyers never notice. The model table has a Standard tier and a Fast tier, and the Fast tier roughly doubles the per-minute rate on the same model - GPT 5.5 is $0.16 standard and $0.32 fast as of August 2026. Fast is bought for latency, which is the right reason, but it is the most expensive way to buy latency and it is worth testing whether a smaller standard-tier model gets you there for less.
The other thing to watch is that the language model rate is per minute of call, not per token. That is convenient for forecasting and it means a caller who sits in silence still accrues model cost. Keeping calls tight is a cost strategy here in a way it is not on token-billed platforms.
Bottom line: The four small per-minute add-ons a real deployment ends up enabling add $0.025 a minute, which is $100 a month at 4,000 minutes - more than the language model costs on a lean configuration.
What breaks in production, and how you find out
Retell documents its concurrency behaviour in more operational detail than any competitor here, and the detail matters because it describes exactly how a caller experiences your capacity limit. Pay-as-you-go workspaces get 20 concurrent calls by default. When an inbound call arrives and no slot is free, it waits. If no slot opens after about 40 seconds, Retell either transfers the caller to the `fallback_number` configured on that phone number, or - if there is no fallback, or the transfer fails - ends the call with a `concurrency_limit_reached` status. Outbound calls without burst enabled are rejected outright.
Read that again as a customer experience rather than an API contract: without a fallback number configured, your busiest hour produces callers who wait forty seconds and then get hung up on. Configuring `fallback_number` to your actual reception line or a voicemail is a five-minute job that converts your worst failure mode into an ordinary one, and it is the first thing to set up on any Retell deployment. Then alert on `concurrency_limit_reached`, because that status is the only signal you will get.
For genuine peaks, Retell offers concurrency burst, which temporarily exceeds the standard limit at a $0.10 a minute surcharge. That is a sensible safety valve and roughly doubles your per-minute cost while it is active - fine as insurance against a busy Monday, expensive as a permanent operating mode. Concurrency and calls-per-second limits are both viewable and adjustable from Settings then Limits in the dashboard, so you can raise the floor rather than relying on burst.
The generic voice failure modes apply as everywhere: latency creeping to the point where the caller talks over the agent, mishearing of names and postcodes, and an escalation path to a human that nobody tested. Test the handoff to a person harder than you test the happy path, because it is what protects you when everything else fails.
Concurrency and compliance, the two ceilings
Twenty included concurrent calls is the most generous default in this comparison and it is the reason Retell suits businesses with an actual phone queue. Raising it is $8 per concurrency per month, so taking a clinic from 20 to 40 lines is $160 a month - a real number, but a published one, and cheaper per line than burst at $0.10 a minute if you exceed the ceiling regularly rather than occasionally.
The compliance ceiling is harder. A signed BAA and custom MSA or DPA terms sit on Retell's Enterprise plan as of August 2026, which means any dental practice, medical clinic or veterinary practice handling protected health information cannot simply sign up and start. PII removal at $0.01 a minute is a useful control but it is not a BAA and should not be treated as one. Price the Enterprise conversation before the build, not after, and get call recording and consent handling written into the same agreement - recording rules differ by state and country and are your obligation, not the platform's.
Bottom line: Set `fallback_number` before launch: without it, a caller who arrives when all 20 concurrent slots are busy waits forty seconds and is then hung up on.
What it connects to that you already run
Retell's telephony is built on Twilio for US numbers at published rates, which is a practical advantage: Twilio is the incumbent most businesses with any existing telephony automation already have an account with, and number porting into a Twilio-backed platform is well-trodden ground. Rates vary by country, so check yours rather than assuming the US figure.
On the AI side, the component model means you are choosing from a menu of providers rather than accepting a vendor's bundle. Text-to-speech options published as of August 2026 include Retell's own voices alongside Minimax, Fish, Cartesia and OpenAI at $0.015 a minute, with ElevenLabs at $0.040. Language models span the current OpenAI, Anthropic and Google line-ups at published per-minute rates, and a custom LLM option is listed for teams that want to bring their own. Being able to switch a provider because it sounded wrong on a real call, without changing platforms, is worth a lot during the first month.
For business systems, the pattern is the standard one for developer-oriented voice platforms: the agent calls functions mid-conversation and fires webhooks at call end, and those endpoints are yours. In practice most small businesses point them at whatever automation platform they already run, letting an existing workflow do the CRM write or the calendar booking. Retell's knowledge base feature covers the read side - opening hours, services, policies - without needing an integration at all, at $0.005 a minute.
Bottom line: Retell charges nothing for telephony if you bring your own SIP trunk, which removes $0.015 a minute - $60 a month at 4,000 minutes - and the largest switching cost in voice at the same time.
Retell AI alternatives: what to pick instead, and when
Retell's proposition is that it publishes everything, so the alternatives divide neatly by what they refuse to publish and what they bundle instead. If you find yourself wanting a Retell alternative, it is almost always for one of three reasons: the itemised bill is too fiddly, the BAA is stuck behind Enterprise, or you want somebody else to choose the components.
Pick Bland AI if the itemisation is the problem. One rate covers the model, transcription and premium voices, and at 4,000 minutes the Start plan is $560 with no platform fee against Retell's $354 to $1,082. You are paying a premium at the lean end for never having to think about a dropdown again, and for many businesses that is a fair trade.
Pick Vapi if you want to go further in the other direction and hold your own provider accounts, so a model price change hits your account rather than being repriced by a platform. Vapi's own layer is $0.05 a minute against Retell's $0.055, but it includes 10 concurrent calls against Retell's 20 and publishes none of the component rates, so the forecast is on you.
Pick ElevenLabs Agents if voice quality is what is actually deciding this - Retell will happily sell you ElevenLabs speech at $0.040 a minute, but the native product is a flat $0.08 a minute with the voice library attached. Pick Phonely if concurrency is your constraint rather than cost, since it is unlimited on every paid plan. Pick Synthflow only at contact-centre volume, where its $30,000 annual floor stops being absurd. And if you are not sure the phone is your bottleneck at all, a free automation audit is a cheaper way to find out than a pilot.
| Platform | Billing unit | At 4,000 minutes | Who it suits | Does better than Retell AI |
|---|---|---|---|---|
| Retell AI | Itemised per-minute components | $354 lean to $1,082 premium | Teams that will tune the model choice | Publishes every component rate; 20 concurrent calls free |
| Bland AI | One bundled per-minute rate | $560 on Start, no platform fee | Businesses that want a single number on the invoice | Nothing to configure; no dropdown can double your bill |
| Vapi | $0.05/min platform, providers at cost | $235 verifiable, plus 3 AI components | Teams holding their own provider keys | Model and voice spend lands in your own accounts |
| ElevenLabs Agents | Subscription minutes at a flat $0.08 | $320 plus model and telephony | Businesses where the caller's first impression is the product | Best synthesised voice, and a rate that never changes with volume |
| Phonely | Subscription minutes with high overage | ~$1,125 on Pro at $0.30 overage | Bursty call patterns that breach concurrency limits | Unlimited concurrent calls on every paid plan |
| Synthflow | Enterprise annual contract | $2,500/mo floor, no per-minute rate published | Contact centres with legacy telephony to integrate | Launch support and scoping included in the contract |
Bottom line: Retell AI is the right pick when somebody will actually tune the model choice; if nobody will, Bland AI's flat $560 at this volume beats a badly configured Retell agent by a wide margin.
Migration and lock-in: can you get your work out?
Retell's commercial lock-in is close to nil, and that deserves credit. Pay-as-you-go with no minimums and no contract means leaving costs you nothing but the rebuild, and $10 of free credit means evaluating costs nothing at all. There is no annual commitment funnelling you toward a renewal date, which is a genuinely different posture from the enterprise end of this category.
The technical lock-in is likewise mild because of the component model. Your model provider, your voice provider and your telephony carrier are all named third parties rather than Retell-internal black boxes, so the pieces doing the actual work are replaceable and, in the case of Twilio-backed numbers, portable through a well-understood process. Bringing your own number rather than provisioning a $2-a-month Retell one removes the single largest switching cost entirely, and it is worth doing on that basis alone.
What you rebuild on a move is Retell's layer: the agent configuration, the prompt and conversation design, the function definitions, the knowledge base contents and the webhook wiring. The knowledge base is the piece worth protecting deliberately - keep the source of your opening hours, service list and policies in a document you own and treat the uploaded version as a copy, so a move is a re-upload rather than a rewrite.
The conversation design is, as always, the asset. Months of refinement into what the agent says when a caller is angry, how it confirms a spelling, when it gives up and transfers - that is the valuable artefact and no platform will import it. Keep it in a document outside the vendor and the rebuild is a week rather than a quarter.
Bottom line: Retell AI has no contract and no minimum, so the only thing binding you is the phone number - and bringing your own removes even that.
Retell AI FAQ
What will Retell cost us per month?
Work it from minutes. On the reference workload of 500 calls a month at three minutes each - 1,500 minutes - the published range of $0.07 to $0.31 a minute as of August 2026 puts you between $105 and $465. That spread is almost entirely your language model choice, which ranges from $0.003 to $0.16 a minute. Add $2 a month for a phone number, and per-minute extras only if you need them: knowledge base, denoising and guardrails are $0.005 each.
Which components actually drive the cost?
The language model, by a distance. Voice infrastructure is fixed at $0.055 a minute, text-to-speech at $0.015 and US telephony at $0.015 - that is $0.085 you will pay regardless. The model swings from $0.003 to $0.16, which is more than the rest of the stack combined at the top end. So the single highest-leverage decision you make is which model to run, and the honest way to decide is to test the cheap one against your real calls before assuming you need the expensive one.
What breaks with Retell in production?
The same things that break every voice agent: latency that makes the caller think the line has dropped, transcription failing on accents and background noise, and the agent not recognising a caller who wants a human. Retell's specific gotcha is cost creep from optional per-minute features - denoising, guardrails, PII removal and especially AI quality assurance at $0.10 a minute - each of which is individually small and collectively significant. Audit what is enabled monthly.
Who is Retell wrong for?
Businesses that need one predictable number on a budget line, because the itemised model produces a bill that moves with usage across a dozen dimensions. Also international operations without doing the telephony homework, since the published $0.015 a minute is a US Twilio rate and other countries differ substantially. And, as with all voice AI, any business whose calls genuinely need human judgement rather than information handling.
Is 20 concurrent calls enough?
For most small businesses, comfortably. Concurrency is about simultaneous calls, and 20 covers a very busy period for a business handling a few hundred calls a month. Extra concurrency is $8 per call per month above the 20 free as of August 2026, which is cheap insurance if your volume is spiky - a Monday-morning rush or a campaign response. Model your peak hour rather than your monthly total, because the peak is what determines whether callers get through.
How locked in are we?
Moderately, and less than with bundled providers. Your prompts and conversation flows are portable text, phone numbers can generally be ported to another provider, and the underlying model and voice providers are industry-standard rather than proprietary. What does not travel is the agent configuration and the call history. Export transcripts periodically - they are both your best test set for any replacement and a genuinely useful record of what customers actually ask.
Should we turn on AI quality assurance?
For the first hundred minutes, definitely - they are free, and finding out how your agent actually performs on real calls is worth more than any amount of pre-launch testing. Beyond that it is $0.10 a minute, which on a cheap configuration nearly doubles your cost. The sensible pattern is to run it on a sample rather than everything: enough calls to spot patterns, not so many that the monitoring costs more than the agent.
Retell or Bland for a small business?
Bland bundles the model, transcription and voice into one per-minute rate, which is simpler and starts at $0.14 a minute with no platform fee as of August 2026. Retell unbundles and can go as low as $0.07 all-in if you choose a cheap model. So Retell is cheaper when you optimise and more complicated to reason about; Bland is more expensive at the low end and simpler. If nobody will spend time tuning, Bland's simplicity is worth the difference.
Can you set up a phone agent for us?
Yes, and the part that makes the difference is not the platform - it is deciding what the agent should do, what it must never do, and when it hands off to a person. We scope that on a free automation audit, build it on whichever provider fits your volume and compliance needs, tune the model choice so you are not paying flagship rates for booking a haircut, and hand it over with prompts, number and documentation you own plus a recurring report of what it handled.
Looking at alternatives? Retell's itemised pricing is the most buyer-friendly thing in this category - you can build an accurate forecast from the published page alone, which is not true of most competitors. Use that: pick the cheapest model that passes a real test with your own calls, and only add the per-minute features you can justify. On lock-in, your prompts and call flows are portable and phone numbers can be ported, so switching cost is moderate rather than severe. The main thing to get right is the model choice, because it is the single largest variable in your bill. See the Autonoly review.
