A voice agent bills on four meters at once

The standard architecture is a cascade. Caller audio streams to a speech-to-text model, the transcript goes to an LLM, the LLM’s text goes to a text-to-speech model, and the resulting audio streams back down a phone line. Four systems, four vendors, four billing units. Every layer streams into the next rather than waiting for completion, because the whole chain has to fit in a conversational gap.

That gap is tight. Platform builders target roughly 50–100ms per phase and a total voice-to-voice budget under 500–700ms — the point past which a caller starts to perceive a pause rather than a response. The architecture is shaped by that constraint, and so is the cost, because the choices that reduce latency and the choices that reduce cost are frequently the same choices.

Here is the mismatch that makes voice-agent budgeting harder than text-agent budgeting. Transcription bills per minute of audio. The model bills per million tokens, split between input and output. Text-to-speech bills per character spoken. Telephony bills per minute of connected call. The orchestration platform on top typically bills a flat per-minute fee.

Only two of those scale with wall-clock time. The other two scale with how much text moves through the system, which is a function of your configuration and the shape of the conversation. A ten-minute call where the agent barely speaks and a ten-minute call where it monologues cost the same on transcription and telephony, and can differ by several multiples on TTS and LLM.

The first non-obvious line item: a voice agent transcribes the assistant as well as the caller. Both channels are billed. That means the transcription meter runs at roughly 2× the wall-clock minutes of the call, not 1×.

It is also the least controllable of the four. You cannot shorten it with prompt engineering — it scales purely with call duration. The only real levers are picking a cheaper transcriber and keeping calls short, which is why the cheapest optimization in voice is almost always ending the call sooner, since it discounts every meter simultaneously.

Text-to-speech bills characters. This makes prose style a cost decision in a way that has no analogue in text agents, where nobody pays extra for a chatty response beyond its output tokens. Here you pay twice: once in output tokens from the model, again in characters at the voice provider, and a third time indirectly because a longer utterance occupies more wall-clock minutes on the transcription and telephony meters.

Premium voices compound this. Component estimates commonly put TTS around $0.04 per minute at premium tiers against roughly $0.01 for transcription — so the voice can cost four times the ears. Those component rates are third-party estimates rather than published provider figures; substitute your own vendor rates before treating the ratio as authoritative.

Now the meter that dominates. LLM input cost is input tokens per minute × effective input rate, and the first term is where voice diverges sharply from text.

A conversational turn is not one model call. Tool invocations, retries, structured-output extraction, and intermediate reasoning steps all issue their own requests, and every request re-sends the full system prompt and tool definitions. Vapi’s published cost model assumes 5 requests per turn — and each of those five carries your entire prompt.

Put a number on it. At 4 conversational turns a minute — an assumption, not a published figure, so vary it against your own call data — that is 20 requests per minute. A 3,000-token system prompt plus tool schema therefore ships 20 × 3,000 = 60,000 input tokens every minute of conversation. Your prompt is billed roughly twenty times per minute, and it is billed whether or not anyone said anything interesting.

Continue the example with commodity mid-tier rates of $1 per million input and $5per million output. Prompt caching matters here: a large stable prefix qualifies for a discounted cached rate, and Vapi’s model blends standard and cached rates at an assumed 50% hit rate. Take the cached rate at 10% of standard — my assumption, stated so you can change it — and the effective input rate becomes (0.5 × $1.00) + (0.5 × $0.10) = $0.55 per million.

Input cost is then 60,000 × $0.55/M ≈ $0.033 per minute. Output is small by comparison: production data puts generated speech around 150 output tokens per minute, so 150 × $5/M ≈ $0.00075. Input outweighs output by more than 40×.

That ratio is the whole lesson. In a chat product output tokens are usually the expensive half. In a voice agent the model barely writes anything — it writes a sentence or two per minute — but it reads your entire prompt twenty times. Optimizing output length is nearly pointless; optimizing prompt length is where the money is.

Run the sensitivity. Adding 1,000 tokens to your system prompt — a few extra examples, one more tool schema, a paragraph of politeness instructions — adds 20 × 1,000 = 20,000 input tokens per minute, or 20,000 × $0.55/M = $0.011 per minute.

Against a typical $0.05 platform fee, that is a 22% increase in the orchestration line from one paragraph of prompt. Across 50,000 minutes a month it is $550. Teams routinely spend weeks negotiating telephony rates worth a fraction of that while treating the system prompt as free text.

Any published per-minute figure rests on assumptions that break in opposite directions. The assumed 50% cache hit rate is conservative for large stable prompts, where real hit rates run higher and actual cost lands below the estimate. Meanwhile the estimate typically ignores growing conversation history, and history is re-sent with every request — so on long calls actual cost lands above it.

The second effect is quadratic-ish and underappreciated. Minute ten of a call carries nine minutes of accumulated transcript on every one of its twenty requests. Two agents on identical configuration can diverge substantially in cost per minute purely on call-length distribution, which is why a blended average from short test calls will understate a production bill built on long ones.

The measurement story mirrors the cost story. Component latencies are usually reported as medians — time for the transcriber to produce text, time to first token from the model, time to first audio from the voice provider — and a displayed total is often just those three medians summed.

That sum is a floor, not an estimate. It excludes endpointing (deciding the caller has actually finished speaking, which is a judgment call and frequently the largest single contributor) and it excludes network transport. Medians also hide the tail, and voice quality is a tail property: callers do not notice the median response, they notice the one that took two seconds. Compare p95, not p50, when the number matters.

One more constraint shapes every configuration decision. Voice agents run their LLMs with reasoning disabled, because reasoning tokens are generated before the first output token and the latency budget cannot absorb them. That means published reasoning-on benchmark scores do not describe the model you are actually running.

On a reasoning-off basis, even the strongest voice configurations land in a narrow band — Vapi puts the smartest around 20–30 on the Artificial Analysis intelligence index and explicitly says to read the scores relatively rather than absolutely. The practical implication: model choice buys you less capability spread in voice than in text, so the marginal return on paying for a frontier model is smaller here, while the marginal cost is the same.

The alternative topology is a single multimodal model that ingests audio and emits audio, skipping the transcribe and synthesize steps entirely. It removes two network hops from the latency budget and two vendors from the bill, replacing four meters with one audio-token meter.

The trade is control. A cascading pipeline lets you swap any component independently, inspect the transcript at each boundary, apply a pronunciation dictionary, or fail over the transcriber mid-call. Speech-to-speech gives you a black box with better latency and worse observability, and audio tokens are typically priced well above text tokens — so it is not automatically cheaper, only simpler. Audio-token pricing varies enough by provider that it is worth pricing your own workload rather than assuming either direction.

Budget a voice agent by asking which meter each decision moves. Call duration hits all four. Prompt size hits only the model meter, but hits it twenty times a minute. Verbosity hits three. Model tier hits one and buys less than the benchmark headline suggests, because you are running with reasoning off.

The platform fee — the number on the pricing page — is usually the smallest and by far the most predictable line on the invoice. The expensive decision is the one you make in a text editor, writing the system prompt.