AI CostBase

Voice agent cost

A voice agent is four paid services stacked on a phone call: listen (STT) → think (LLM) → speak (TTS) → the line itself (telephony). The LLM is often less than half the bill.

Turns = how many times the agent thinks per call. More agentic = more turns.

The formula

minutes = calls/day × call_length × 30 llm = calls × turns × tokIn/1M × P_in + calls × turns × tokOut/1M × P_out stt = minutes × STT_rate tts = minutes × TTS_rate tel = minutes × tel_rate per_call = (llm + stt + tts + tel) / (calls × 30)

Defaults reflect mainstream hosted APIs as of : Whisper-class STT, mid-tier neural TTS (~450 chars/min), typical US inbound voice-API rate. Token prices update daily from OpenRouter's API. Barge-in (interrupted speech) typically cuts billed minutes 10–20% — not modeled; treat the result as the conservative case.

What actually drives the per-minute number

Four meters run at once on every call, and they do not move together. Knowing which one your workload stresses is the difference between a $0.04/min agent and a $0.30/min one.

MeterTypical rateScales with
Speech-to-text$0.004–0.006 / minCall duration, not turns
LLM turns$0.005–0.060 / minTurns × context size × model tier
Text-to-speech$0.010–0.030 / minCall duration; premium voices cost 3–10× mid-tier
Telephony$0.007–0.020 / minCall duration; inbound US is the cheapest case

Two consequences follow. First, three of the four meters bill on minutes, not on intelligence — so making the model cheaper buys you less than you expect. Second, the model is the only meter you can move by editing code, which is precisely why teams over-optimise it and under-budget the rest.

Why the transcript, not the model, is the real turn meter

Every turn re-sends the running transcript. Turn one might carry 300 tokens; turn eight carries the entire preceding exchange plus the system prompt. On a six-turn call the input side is therefore not 6 × 600 — it is closer to the triangular sum of a transcript that keeps growing, which is why the context-per-turn default here describes a mid-call turn rather than the first one.

The practical effect is counter-intuitive: cutting a bloated system prompt from 1,200 tokens to 400 often saves more than switching the answering model from a frontier tier down to a mini, and it costs you nothing in answer quality on the short transactional intents that make up most call volume.

Where teams get this wrong

Cutting the bill without hurting the call

Concurrency: the cost that never appears on a per-minute rate card

Everything on this page is priced per minute, which describes the marginal call and says nothing about the busiest one. Production voice platforms charge for capacity somewhere — a per-concurrent-session fee, a tiered package, or a committed minimum. An agent answering 100 calls spread across a day needs very little capacity; the same 100 calls arriving inside a two-hour window needs an order of magnitude more.

Before committing to infrastructure, take the per-call figure this calculator produces and multiply it by your expected peak calls per minute rather than the daily average. The two numbers differ by far more than teams expect, and it is the second one that gets priced.

The same distinction applies to the model. A per-token rate is a marginal price; rate limits and provisioned throughput tiers are capacity prices. An agent that performs perfectly in testing and fails at nine o'clock on a Monday morning is not experiencing a model problem — it is experiencing a capacity problem, and capacity is where voice projects most often discover a cost they never modelled.