An AI voice agent usually costs between $0.05 and $0.20 per minute, and the reason it’s higher than a text chatbot is that every minute pays three bills at once: speech-to-text to hear the caller, an LLM to decide what to say, and text-to-speech to say it. Add telephony on top. At the middle of that range, 1,000 five-minute calls runs roughly $500 to $750. The voice agent cost calculator adds up all three services plus telephony so you see the real per-minute and per-call number, not just the LLM slice.
Three services, one conversation
A text chatbot has one cost: the LLM. A voice agent has three, running in sequence on every turn:
- Speech-to-text (STT). The caller’s audio is transcribed into text. Billed per minute of audio, usually the cheapest of the three — often around $0.006 to $0.02 per minute.
- The LLM. The transcript goes to a language model that decides the reply. Billed per token, input and output, and it resends the growing conversation each turn.
- Text-to-speech (TTS). The reply text is turned into spoken audio. Billed per character or per minute of speech, and premium natural voices cost several times more than basic ones.
Miss any one of these and your estimate is wrong. Miss telephony — the cost of the actual phone line or SIP connection — and you’re wrong again. All four stack into the per-minute figure. These prices move often and vary by vendor, so treat the numbers here as approximate starting points and plug your own provider’s rates into the calculator.
Where the cost concentrates
The three bills are rarely equal. In most setups:
- TTS is often the biggest line, especially with premium, human-sounding voices. Cheap robotic voices are far less, but the natural ones people actually want are the priciest per minute.
- The LLM is next, and it grows with conversation length because each turn resends the whole transcript. A long, chatty call costs disproportionately more than a short one — the same compounding that makes text agents expensive.
- STT is usually the cheapest, a small and fairly fixed cost per minute of audio.
- Telephony is separate and easy to forget — a per-minute carrier charge that’s small individually but real at volume.
Knowing which line dominates tells you where to optimise. If TTS is 60% of your bill, a cheaper voice saves more than any LLM tweak.
A worked example: a support line
Say you’re running an AI support line. An average call lasts 5 minutes, with the caller and agent talking about half the time each.
- STT: 5 minutes of caller audio at ~$0.01/min = $0.05.
- LLM: roughly 10 back-and-forth turns. With the conversation resent each turn, call it ~15,000 input and ~2,000 output tokens total. At $3/M input and $15/M output: $0.045 + $0.03 = $0.075.
- TTS: ~2.5 minutes of spoken replies. At a premium voice around $0.06/min = $0.15.
- Telephony: ~$0.01/min × 5 = $0.05.
Per call: ~$0.325. Per minute: ~$0.065. Across 1,000 calls that’s about $325 — and note the TTS line ($0.15) is nearly half of it, with the LLM second. Swap to a cheaper voice and the whole per-call cost drops faster than any other single change. Lengthen the calls and the LLM line climbs, because of the resend-every-turn effect.
Cost per minute and per 1,000 calls
To get from one call to a budget:
- Cost per minute = total call cost ÷ call length. Useful for comparing setups regardless of how long calls run.
- Cost per 1,000 calls = per-call cost × 1,000. This is the number that matters for capacity planning — it turns an abstract per-minute rate into a line item you can forecast.
Both fall straight out of the four-part estimate. The voice agent cost calculator lets you set each service’s rate, your average call length and turn count, and reports the per-minute and per-1,000-call totals so you can test a cheaper voice or a smaller LLM before committing.
The compounding you can’t ignore
The LLM line grows with conversation length because voice agents, like all agents, resend the full transcript on every turn. A call that drifts from 5 minutes to 10 doesn’t just double the audio bills — it more than doubles the LLM cost, because later turns carry a longer history. Why AI agents cost so much explains that compounding in depth, and it’s the single biggest reason voice budgets overshoot. If latency matters too — long pauses make an agent feel broken — tokens per second explained covers the speed side of picking a model.
