Retell AI Pricing Explained: $0.07 Floor, $0.11 Default

Retell's pricing page advertises $0.07 to $0.31 a minute. That is a 4.4x spread inside one product, and it is the honest way to quote it.

The trouble is what the floor is made of. $0.07 is Retell Voice Infra at $0.055 plus the cheapest platform voice at $0.015, and nothing else. It is reachable only by bringing your own language model and your own carrier and paying both of them outside Retell's invoice.

Drive the vendor's own calculator and it says as much without being asked. At its untouched defaults it returns $0.11 a minute, itemised as $0.04 of model, $0.055 of infrastructure and $0.015 of voice. Set every menu to its dearest option and it returns $0.27.

So there are three numbers on one page, and only one of them is the one people quote.

What follows: what the $0.07 is made of, two different all-in figures computed on two different bases, the dropdown that moves your bill more than the platform choice did, one supported option that costs more than the advertised ceiling, the add-ons that make an unattended agent defensible, and the correction this site had to publish about its own Retell arithmetic. For the head-to-head against the other two platforms people shortlist alongside it, that is a separate page: Retell vs Vapi vs Bland.

What does Retell AI cost?

What 'all-in' means here, and where the component floor comes from

Two figures on this page look like the same kind of thing and are not, so it is worth separating them before the arithmetic starts.

An advertised rate is what a platform charges for the parts it supplies. Retell's $0.07 is voice infrastructure at $0.055 plus the cheapest platform voice at $0.015, and it excludes the language model and the telephony. Nothing about that is dishonest; it is a rate card with two rows left off the bottom.

A computed all-in is the advertised rate plus the published list price of every component it excludes. This site holds those component rates in one file so that a correction to any of them corrects every page at once, and none of them is estimated or averaged across vendors.

Retell excludes two of those four, at $0.0108 and $0.0140, which is where $0.07 becomes $0.0948. Read that as a floor and not a forecast: it assumes list pricing, one clean turn per exchange, and no retries, no failed calls and no concurrency minimums. It is still useful, because the gap between the advertised rate and the floor is already running 1.35x on this vendor and up to 3.3x elsewhere in the category [both computed].

$0.0948 a minute is the computed all-in floor, against a $0.07 advertised floor. That is a multiple of 1.35, and it is the smallest gap between advertised and all-in of any bring-your-own-component platform on this site.

The computed figure adds two things the $0.07 leaves out: a low-latency language model at $0.0108 a minute and US telephony at $0.0140, both list rates with verification dates held in this site's cost model. It adds nothing else, because Retell already bundles speech recognition inside the infrastructure line and synthesis inside the voice line.

It is a floor rather than a forecast. It assumes list pricing, no committed-use discount, and one clean turn per exchange. Real invoices run higher.

What the $0.07 is actually made of

Two line items, both Retell's, and it decomposes exactly on the vendor's own component table.

Nothing here is hidden. Every one of those rows is printed on Retell's pricing page, which is more than almost anybody in this category does. What is missing is the addition, and three of Retell's own comparison pages still call the $0.07 flat.

The advertised rate is a floor, not a rate. Read it as the price of the two components Retell insists on providing, and treat the other two as separate purchasing decisions you have not made yet.

Two all-in numbers, and both are built from Retell's own rates

This is where the page earns its keep, because the two figures are computed on different bases and mixing them is how people end up arguing past each other.

The first uses this site's component cost model: mid-market list rates for the things Retell excludes, chosen once and applied identically to every platform so the comparison is like for like. That gives $0.0948.

The second uses Retell's own rate card and nothing else. Take voice infra at $0.055, the cheapest platform voice at $0.015, Retell's own telephony at $0.015, and add a model from Retell's own table.

Both numbers are defensible and they answer different questions. The $0.0948 answers whether Retell is expensive relative to its competitors on a common basis. The $0.130 to $0.245 range answers what your invoice will look like if you buy everything from Retell, which is what most first deployments do.

One dropdown decides your bill

Retell publishes a per-minute price for every model it offers, which almost nobody in this category does. Twenty-plus of them, each with a separate faster tier at roughly double the rate.

That is a fifty-fold spread on one menu, sitting inside a product whose headline rate people argue about in tenths of a cent.

It is also the good news. A team that picks a fast cheap model deliberately, rather than accepting whatever the calculator defaults to, saves more than it would by switching platforms. This site makes the same argument about synthesis in the component audit: the dropdowns move the total further than the vendor choice does, and nobody asks about them in a demo.

One supported option costs more than the advertised ceiling

The page quotes $0.07 to $0.31 a minute. It also lists GPT Realtime as an alternative speech-to-speech engine at $0.345 a minute, for that layer alone, before telephony or add-ons.

A buyer who selects it is above the top of the advertised range on one line item.

This is not a gotcha so much as a limit of the format. A range implies a bounded product, and a rate card with twenty models and seven add-ons on it does not have an upper bound that a two-number range can express.

The add-ons that make an unattended agent defensible

Read these as a group rather than individually, because they are the lines that let you put an agent on the phone without a human watching, and every one of them is an extra.

Quality assurance alone costs more than the advertised floor of the whole product. Turn on all four and you have added $0.12 a minute to a rate advertised at $0.07.

Whether you need them is a real question rather than a rhetorical one. Guardrails and PII removal are the two most likely to be non-negotiable if you are calling consumers or handling anything regulated, and they should be in your model from the first day rather than discovered in month three.

What Retell costs an hour, which is the unit buyers actually compare

The ten platform rates this directory holds, with what each one excludes

Retell's rate card is unusually complete and it is still one vendor's. Set beside every other per-minute platform this site holds a rate for, with the excluded components named and the all-in computed on the same four component prices, the ranking by advertised rate and the ranking by payable rate are not the same ranking.

Why the model column decides the ranking

Retell is fifth on advertised rate and third on computed all-in, which is a smaller move than several rows around it make. Millis advertises $0.02 and pays $0.0665 because it bundles nothing at all. Ultravox advertises two and a half times Millis and pays $0.0640, the cheapest in the table, because it bundles everything except the phone call. Ranking this category on headline rates puts those two in the wrong order.

The column that matters most on this page is the third. Retell excludes the model, and the model is the dropdown carrying a 53x spread from $0.003 to $0.16 a minute on Retell's own published table. So the $0.0948 in the fourth column is computed on this site's mid-tier reference model rather than on whichever one you pick, and a Retell agent on GPT 5.5 computes to $0.245 a minute [computed], which is above every other row here including Bland's $499-a-month tier.

Per-minute pricing hides the comparison people are really making, which is against a person.

Someone made this conversion on Retell's own Launch HN thread on 22 February 2024, item 39453402, 350 points and 173 comments. User reissbaker wrote: "I do wish Retell's pricing was cheaper, though; $6/hr is pretty much the cost of a call center employee in India, and LLMs still perform below the average human on most things."

In the same thread, user nostrebored ran the substitution against his own contact-centre baseline: "If I were to bring retell into the loop, I'm changing my self-service per minute cost from .018 + .004 = .0184 per minute to .1184 per minute at the cheapest setting." That is 6.4 times, computed by a buyer rather than by a vendor, on what he took to be the cheapest configuration.

Both are one person characterising their own stack in early 2024 rather than a finding of fact, and the rates have moved since. They are quoted because they are the only public arithmetic on this product done by somebody with no stake in the answer, and because the unit they reached for is the one the pricing page will not print.

Chat and SMS are metered on a different unit

Retell's chat agents are priced per AI message rather than per minute, from $0.001 on GPT 5 nano to $0.052 on GPT 5.5. SMS is a flat $0.01 a message on top.

A per-message meter and a per-minute meter cannot be compared without fixing a workload, and Retell does not publish a conversion between them. If you are sizing both channels, count messages and minutes separately and price them separately, because the intuition that transfers between them is wrong in both directions.

Credit where it is due: Retell publishes the rate card this site normally builds

This site exists because voice platforms advertise an orchestration rate and stay quiet about what it excludes. Retell does the opposite.

Voice infrastructure, every text-to-speech voice, every model, telephony and each add-on all carry their own per-minute price on the pricing page. The exclusions are stated. The range is quoted as a range. The calculator itemises its own output.

That is more disclosure than any other platform in this category offers, and it is the reason this page can compute a second all-in figure from Retell's numbers alone. The criticism that remains is narrow: the page performs none of the addition, and three of Retell's own comparison pages describe the floor as flat when the page itself no longer does.

This site had Retell's multiple wrong

The correction generalises. An exclusion list is the input that decides everything downstream, and it is the one field most likely to be filled in from a summary rather than from the page. Read the exclusion sentence before the price sentence, every time.

How to price your own Retell minute in ten minutes

• Pick the model first, not last. It is a fifty-fold spread and it is the largest single lever on the bill. Decide whether you need the frontier tier before you compare platforms at all.
• Decide who carries the phone call. Retell's own telephony is $0.015 a minute. Bringing your own carrier moves that line onto a different invoice rather than removing it.
• Add the compliance add-ons you cannot decline. Guardrails and PII removal are $0.015 a minute together, and for regulated calling they are not optional.
• Run the vendor's calculator at your own settings and write down what it returns. If your number is far from its default $0.11, you have found the assumption that matters.
• Convert to an hour and to a month before you take it to anyone who signs things. Per-minute rates are unreadable as budget lines.

When Retell is the wrong purchase

What breaks in month three

The model bill moves without anyone changing anything. Retell prices every model separately and model prices move faster than platform prices, so the line item that carries a 53x spread is also the line item most likely to be repriced by somebody other than Retell. A budget built on $0.045 for GPT 4.1 is a budget with a third-party dependency in it, and nothing on the invoice will announce the change.

The add-ons that make an unattended agent defensible get switched on after launch, not before. PII Removal at $0.01 a minute is a compliance requirement rather than a preference for most regulated deployments, and AI Quality Assurance is $0.10 a minute after the first hundred minutes, which is more than the advertised platform rate. Both usually arrive in month two or three, when somebody senior listens to a call.

And the retries are not in any published figure. Every rate on this page is per connected minute. Failed calls, silence timeouts and transfer legs are real traffic that does not appear in a rate card, and the gap between a computed floor and an invoice is mostly made of them. That is the measurement named at the end of this page, and it is the one nobody has run.

When you are at volume and will genuinely do component optimisation. The bundle's advantage is time rather than money, and a team that will actually swap in cheaper components as they ship gets more out of a bring-your-own-key platform like Vapi. The catch is in the word actually. Teams that intend to and never get round to it pay more than they would have with the bundle.

When you need a specific voice or model the bundle does not carry, or on-premise deployment, which it does not offer.

And when swapping components matters to your economics. You cannot replace a layer when something better ships; you wait for the vendor. In a category moving this fast that is the strongest argument against bundling regardless of what the arithmetic says today.

What this page does not know