How to Price an AI Voice Agent

Voice agents have the messiest bill in AI: three metered services stacked on every single minute of conversation, plus a telephony layer nobody remembers until the invoice arrives. This guide walks the full math — what a minute actually costs you with live prices, what quietly inflates it, and how builders turn that cost into a price a client will sign. Prefer to plug in your own numbers? The voice AI cost calculator does this interactively.

A voice minute is three bills, not one

Every conversation minute passes through speech-to-text (hearing), an LLM (thinking), and text-to-speech (speaking). Each is metered differently — STT and TTS per minute of audio, the LLM per token — which is why most estimates only price the layer they thought of first:

• Speech-to-text: $0.0043–$0.016/min (Deepgram Nova is the current floor, checked 2026-08-02).
• LLM reasoning: a turn-by-turn agent burns ~1,800 input + ~220 output tokens per minute, because context is re-sent every turn. On OpenAI GPT-5.4 Mini that is ~$0.0023/min; on OpenAI GPT-5.6 Sol it jumps to ~$0.0058/min (live prices, updated 2026-08-23).
• Text-to-speech: $0.015/min for standard voices (OpenAI TTS) up to $0.1/min for premium quality (ElevenLabs (Creator tier)) — a 6–7× spread on the layer your callers actually hear.

What a minute costs: two real stacks

Two things jump out. First, the spread: the premium stack costs 5× the budget one, and almost all of that is the TTS and flagship-LLM choices. Second, even the premium stack sits far below a human call-center minute (~$0.75–$1.50 fully loaded) — the economics work; the question is only how much margin you keep.

The 20–30% that never makes the first estimate

• Telephony: if calls ride phone lines, add ~$0.007–$0.02/min (Twilio-class pricing) on top of the AI stack.
• Silence is billable: STT meters wall-clock audio — hold music, pauses, and dead air transcribe to nothing but still cost money.
• Interruptions: when a caller talks over the agent, the TTS audio you already generated is thrown away and regenerated.
• Logging and QA: transcripts, analytics, and eval runs are small per call and real at scale.

Rule of thumb: multiply the raw stack cost by 1.25 before quoting anyone, including yourself.

From cost to price: what builders actually charge

If you build agents for clients, cost per minute is the floor, not the price. Three models dominate:

1. Per-minute, 2.5–4× cost: the budget stack above lands at ~$0.027/min once padded; billing $0.10–$0.25/min keeps a healthy margin and still reads as 3–7× cheaper than the human agent it replaces. Easiest to sell, easiest to compare — but your revenue shrinks as models get cheaper, so re-anchor quarterly.
2. Per-resolution: a flat fee per completed outcome (booking made, ticket deflected). Anchors against the value of the call, not the cost — typically $0.50–$3 per resolution where a human handling the same call costs $3–$8. Requires solid outcome tracking before you can invoice it.
3. Platform + usage: a monthly fee ($300–$2,000 covering setup, prompts, monitoring) plus a metered rate above an included minute quota. Smooths your revenue and matches how clients budget.

Price your own agent in five steps

1. Pick your stack tiers and get the raw per-minute number from the voice AI cost calculator — try the mini-LLM default first; most routine calls don't need a flagship.
2. Multiply by 1.25 for telephony, silence, and interruptions.
3. Multiply by expected volume (calls/day × avg minutes × 30) — that's your monthly cost of goods. Sanity-check the LLM line against the API cost calculator.
4. Choose a pricing model from the section above; if per-minute, start at 3× padded cost.
5. Re-run the math when prices move — the LLM layer on this page updates 2026-08-23, and model prices drop often enough to widen your margin for free (see the cost-cutting levers for routing and caching tricks that cut the LLM line further).