Part V · Production
Chapter 23
Cost
Five line items that add up to four to eight cents a minute, why speech-to-speech costs several times that, why a human costs many times that, and six ways to spend less.
On this page
In this chapter, we will learn what one minute of a call costs in the cascaded pipeline, line by line, and how that compares with a speech-to-speech model and with a human agent. The prices are the rough public prices the post quoted on 16 September 2026, and the post is clear that actual numbers vary a lot by provider. What stays true is the shape: where the money goes and what makes it smaller.
The five line items
The cost per minute of a cascaded call is the sum of five line items: telephony, STT, LLM, TTS and infrastructure.
| Line item | Per minute | Why |
|---|---|---|
| Telephony | around $0.01 | The phone line, through the provider. |
| STT | $0.005 to $0.01 | Streaming speech-to-text for the whole minute. |
| LLM | $0.01 to $0.03 | A 5-minute call has around 10 to 15 turns, each with a few thousand input tokens and a hundred output tokens. With a small model and prompt caching. |
| TTS | $0.01 to $0.03 | Per minute of generated audio. The agent speaks for around half the call. |
| Infrastructure | around $0.005 | Gateway, orchestrator, storage. |
low $0.01 + $0.005 + $0.01 + $0.01 + $0.005 = $0.04 per minute high $0.01 + $0.01 + $0.03 + $0.03 + $0.005 = $0.085, roughly $0.08 per minute
So the cascaded pipeline costs roughly $0.04 to $0.08 per minute. The LLM and the TTS are the two lines with the widest range, and they are the two we can do most about.
Two comparisons
A speech-to-speech model today costs several times the cascaded pipeline, often $0.10 to $0.30 per minute, because audio tokens are expensive. A human agent costs $0.50 to $1.00 per minute in many markets. Put the three side by side and we can see why companies are investing here.
cascaded pipeline $0.04 to $0.08 per minute speech-to-speech $0.10 to $0.30 per minute several times cascaded human agent $0.50 to $1.00 per minute 6 to 25 times cascaded ($0.50 / $0.08 to $1.00 / $0.04)
Telephony and infrastructure are fixed at the post's figures; the three model lines are sliders that move only inside the post's ranges. Before you move anything, guess what a day of calls costs at the scale from Chapter 4. Then move the LLM slider from its low end to its high end and watch the daily number.
Six ways to reduce cost
- Use small models where possible, and route to big models only when needed.
- Prompt caching for the system prompt and the tools.
- Summarize long contexts, so the input tokens per turn stop growing.
- Do not send silence to STT. VAD decides what goes to STT.
- Self-host open-source STT and TTS models at scale.
- End dead calls quickly.
Five of the six are chapters you have read. Routing and prompt caching are Chapter 9, summarization is Chapter 17, the client-side VAD is Chapter 7, and the dead-call timeout is the first row of Chapter 20. Self-hosting open-source STT and TTS models is new here, and the post adds "at scale": it pays off when there are enough call-minutes to spread the cost of our own GPUs. Cost is not a separate design. It is the same design, read with the invoice in hand.
Try it yourself
In the interactive, set every slider to the top of its range and note the daily cost. Then apply the six savings in your head: which sliders does each one move, and which ones does it leave alone? Notice that telephony and infrastructure are not sliders at all. They are not where the money is.
Common questions
Why is TTS charged for about half a minute per minute of call?
Because TTS is priced per minute of generated audio, and the agent speaks for around half the call. The other half is the user speaking, which costs STT, not TTS.
Are these prices still right?
They are the rough public prices the post quoted on 16 September 2026, and the post says actual numbers vary a lot by provider. Treat the shape as durable and the digits as a snapshot.
If speech-to-speech costs several times more, why would anyone use it?
When the natural feel of the conversation is the product itself, as Chapter 13 said: companions, tutors, casual assistants. For a support line with strict rules, the cascaded pipeline is cheaper and more controllable.
Carry this
- Cascaded: telephony $0.01, STT $0.005 to $0.01, LLM $0.01 to $0.03, TTS $0.01 to $0.03, infrastructure $0.005. Roughly $0.04 to $0.08 per minute.
- Speech-to-speech: $0.10 to $0.30 per minute, several times more, because audio tokens are expensive.
- A human agent: $0.50 to $1.00 per minute in many markets.
- Six savings: small models and routing, prompt caching, summarization, no silence to STT, self-hosted STT and TTS, end dead calls quickly.
Check yourself
1. Which two line items carry most of the range between $0.04 and $0.08?
Answer
The LLM ($0.01 to $0.03) and the TTS ($0.01 to $0.03). Telephony and infrastructure are nearly fixed.
2. Why does a speech-to-speech model cost several times more per minute?
Answer
Because audio tokens are expensive. There are many more of them than text tokens for the same conversation.
3. Name three of the six ways to reduce cost and the chapter each comes from.
Answer
Any three of: small models and routing (Chapter 9), prompt caching (Chapter 9), summarizing long contexts (Chapter 17), not sending silence to STT (Chapter 7), ending dead calls quickly (Chapter 20). Self-hosting STT and TTS at scale is new in this chapter.