Retell AI Review (2026)

Retell bundles its own voice infrastructure and speech recognition into one rate. It does not bundle the language model, and the pricing page has stopped implying otherwise: it now advertises $0.07 to $0.31 a minute rather than a single figure.

The $0.07 floor decomposes exactly, on the vendor's own component table: Retell Voice Infra at $0.055 plus the cheapest platform voice at $0.015. Nothing else. It is reachable only by bringing your own model and your own carrier and paying both of them outside Retell's invoice.

Drive the vendor's own calculator and it says so itself. 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 most expensive option and it returns $0.27.

What bundling actually buys

Not the price. The time.

A bring-your-own-key platform quotes $0.05 and requires you to hold four vendor relationships, tune endpointing across four seams, and diagnose failures that occur between components rather than inside them. Retell quotes more and hands you a working agent.

For a team shipping this quarter, the engineering time saved dwarfs the per-minute difference. For a team at serious volume, the arithmetic inverts, and the question is where.

Where the crossover sits

Take Vapi at $0.05 advertised, which computes to $0.0965 all-in on published component rates. Retell at $0.07 computes to $0.0948. On list pricing they land a fifth of a cent apart, which is a more interesting answer than either headline suggests.

So the honest framing is not bundled versus assembled on price. It is whether you will do the component optimisation. Teams that will, save meaningfully at volume. Teams that intend to and never get round to it pay more than they would have with the bundle.

The trade you are making

You cannot swap a component when something better ships. When a cheaper synthesis model appears, or a faster inference provider, a bring-your-own-key stack absorbs it in an afternoon and a bundled one waits for the vendor.

In a category moving this fast that is not a small thing, and it is the strongest argument against bundling regardless of what the arithmetic says today.

The fastest path to a working phone agent among the developer-facing platforms. Self-serve, free credit, and an agent taking calls the same day without holding four vendor accounts.

• Test interruption handling before anything else. Call your own agent and talk over it. This is where deployments fail and where platforms differ most, and it takes ten minutes to evaluate.
• Check the latency you actually experience, mouth to ear, not the figure in the documentation. The published number is time to first byte and the caller hears something else.
• Add telephony to your model. The advertised rate excludes it, so budget the carrier cost separately from the start.

Published, self-serve, pay as you go, and now quoted as a range: $0.07 to $0.31 a minute. Speech recognition is inside the infrastructure line. The model and the telephony are not, and there are seven per-minute add-ons beyond them.

Computed against this site's cost model, the all-in floor is $0.0948. Set against the alternatives on published rates:

Read that table carefully, because it inverts the usual assumption. The platform with the highest advertised rate has the lowest computed cost, and the one advertising $0.02 computes to nearly the same figure as the one advertising $0.07. Headline rates in this category are close to uninformative.