What a Voice AI Agent Really Costs Per Minute

Five meters run on every minute

The reason per-minute pricing confuses people is that the phrase describes one unit but five products. When a caller is on the line, all of these are counting at once, and only the first is typically what a platform quotes you.

Two consequences follow immediately. First, any quote that consists of a single number has either bundled the other four at a markup, or excluded them. Both are legitimate business models; not knowing which one you are looking at is not. Second, your cost per minute is partly a function of how your agent behaves. A system prompt twice as long, or an agent that answers in three sentences instead of one, moves the bill without anyone changing a price.

The formula, and the minute you actually pay for

The arithmetic is not hard. What trips teams up is the second half: the minute you are billed for is longer than the minute anyone was talking.

cost_per_minute =
      orchestration_fee
    + stt_rate        × audio_minutes_in
    + llm_rate        × (prompt_tokens + output_tokens)
    + tts_rate        × characters_spoken
    + telephony_rate  × billed_legs

billed_minutes =
      talk_time
    + ring_time         (on most carriers, outbound)
    + silence_and_hold  (the caller is still on the line)
    + rounding          (per-second, or up to the next minute)

Ring time is the one that surprises people on outbound. On many carriers the leg starts billing when the call is answered, but campaign economics are dominated by the calls that are not answered, each of which still costs you an attempt. Silence is the one that surprises people on inbound: while a caller thinks, or waits on hold for a transfer, speech recognition may still be streaming and your orchestration meter is certainly still running.

Rounding is the quiet one. Per-second billing and rounding-up-to-the-minute produce very different bills on short calls, and short calls are most calls. A one-line lookup that takes 24 seconds costs either 24 seconds or a full minute depending on a policy that is rarely on the pricing page. On a thousand calls a day, that single detail is the difference between two invoices you would describe differently.

A worked example

Here is a three-minute inbound call on a classic pipeline, with the shape of the spend rather than a promise about anyone's current rates. Provider prices move every quarter, so treat the proportions as the lesson and put today's numbers in the formula yourself.

This is also why comparing platforms on their headline fee alone tells you very little. A platform charging more per minute but letting you bring a cheaper model can easily be cheaper in total, and a platform bundling everything at one attractive rate may be taking a margin on the model calls that you cannot see or shop around.

Where the hidden multipliers live

Five things reliably make real invoices exceed forecasts. None of them are anyone being dishonest; they are all consequences of billing a conversation rather than a request.

1. 01

   The system prompt is charged on every turn

   Conversation history plus the instruction goes back to the model each time the caller speaks. A prompt you doubled in length does not cost double once, it costs double per turn for the rest of the call. Prompt caching, where the provider supports it, changes this arithmetic significantly and is worth asking about explicitly.
2. 02

   A transfer can bill two legs

   Bridging a caller to a human means the platform is holding two calls for as long as the conversation continues. Whether you pay for one or both depends on how the transfer is implemented, and the difference is not visible from the outside.
3. 03

   Outbound retries multiply attempts, not conversations

   A campaign with three retry attempts on no-answer has up to three times the telephony cost per reachable lead. Model your cost per connected conversation, not per dial, or the business case will be wrong by a factor.
4. 04

   Concurrency ceilings turn into lost calls, not higher bills

   This one is the inverse trap. A cap on simultaneous calls does not raise your cost, it silently caps your revenue at peak. Know the number and know what happens to caller number twenty-one.
5. 05

   Recording, storage and analysis are separate decisions

   Post-call analysis runs a model over the transcript, which is another model call per conversation. Storage is small per call and unbounded over a year. Both are easy to switch on and forget.

Platform keys or your own: who gets billed

There are two coherent ways to buy the model layer, and the choice changes your invoice, your margin, and your exposure. Neither is universally right.

With platform keys the provider relationship is your vendor's. You get one bill, you start in minutes, and every supported provider works without you opening an account anywhere. The question to ask is whether the model usage is passed through at cost or marked up, because that single answer determines whether the arrangement is convenience or margin.

With your own keys, you hold the provider accounts and pay them directly. You see the real usage, you can negotiate volume yourself, and you inherit provider-level controls over data handling. You also inherit the failure modes: an expired card or a rate limit on your account is now a live-call problem, and the honest question to ask a vendor is what their platform does at that moment. A platform that quietly swaps to a different provider to keep the call alive has made a decision about your caller's experience that you did not authorise.

Ask both questions in the same breath: is the model usage marked up, and what happens when my own key stops working mid-call.

Realtime models have a different cost shape

A native-audio model that hears and answers in audio replaces three meters with one. That simplifies the arithmetic and usually raises the floor: realtime audio pricing tends to sit above the sum of a lean pipeline, and it is billed on audio duration rather than on tokens and characters, which makes it far more predictable per minute and far less sensitive to how chatty your agent is.

The practical consequence is that the two options have different risk profiles rather than a simple price ranking. A pipeline rewards optimisation, punishes verbosity, and gives you many small levers. Realtime gives you one number that barely moves. For a business modelling a support line at volume, predictability is often worth more than the lowest achievable cost, and for a high-volume outbound campaign the opposite is usually true.

Nine questions to ask before you sign

If a vendor cannot answer these quickly and in writing, that is itself the answer. We publish them because we would rather be compared on them than on a headline rate.