Someone on the finance team pulls up last month's invoice from the voice AI vendor. One line: $0.15/min × 41,860 minutes. No breakdown. Somebody asks the obvious question (is that a good price?) and nobody in the room can answer, because nobody in the room knows what's actually inside that number.
Here's what's inside it, in outline. A "per-minute voice AI call" isn't one meter, it's five or six running at once: an agent-runtime meter, a telephony meter, a language-model meter, a speech-to-text meter, a text-to-speech meter, and usually an observability meter. One invoice, six clocks, and they don't move at the same speed.
LiveKit, who sells the orchestration layer that stitches these pieces together, actually publishes a worked example of this on their own pricing page. It's illustrative, not a plan price, but it's real and vendor-published, and on their numbers the text-to-speech meter alone is very nearly half the entire stack. Not the "AI brain." The voice.
Where $0.0672/min actually goes
LiveKit's own published illustrative example, split into its six components
Source: LiveKit's pricing page (illustrative example, not a fixed plan price).
That's the number people remember, and it's also the one that's easiest to get wrong. LiveKit's own calculator discloses the token volume behind its language-model line, but not which model those tokens are priced on. Swap that one assumption for a model teams actually run in production, and the ranking in this chart doesn't just shift, it inverts: the "AI brain" stops being the rounding error and becomes the largest line on the invoice. Which model you're quietly defaulted into is worth more than every STT and TTS vendor's price difference put together, and it's usually the one line nobody in the room asked about.
None of the other five meters are static either. STT rates differ vendor to vendor, sometimes for real capability differences and sometimes for no reason at all. Prompt caching, when it's actually configured and actually eligible for the model you're on, can cut the LLM line dramatically, and just as easily do nothing if it's silently unavailable on the exact model and session type you're running. Transport is close to a solved problem and rarely where the savings are. TTS is a real cost, but not the mystery it looks like sitting inside a blended rate.
A single blended per-minute rate isn't dishonest. It's just opaque by construction. It collapses six meters with wildly different economics into one number, so you can't tell whether your money is sitting in an expensive voice, an oversized model, or a caching setting nobody ever turned on. The component breakdown shown above is one published vendor example: the split shifts meaningfully with your own model and telephony choices. Working out which of those is actually driving your bill, and which combination of vendor, model, and configuration fits your call volume, isn't a spreadsheet exercise you run once from a pricing page. It's the audit we run before a client signs anything, against their real traffic instead of a vendor's illustrative example.