Quick answer: An AI voice agent minute is never one number — it is the sum of a speech-to-text (STT) line, a language-model (LLM) line, and a text-to-speech (TTS) line, all riding a phone call, and each line prices differently. STT bills per audio minute (~$0.005–$0.02), TTS bills per character (~$0.01–$0.30 per minute of speech), and the LLM bills per token — which makes it the least predictable of the three. Once you time and price each stage separately instead of accepting one blended rate, the math is public: the provider delivers only ~37% of a talk-minute's cost picture, your own agent design sets the other ~63%. This post turns that split into formulas you can rerun on any vendor quote.
1. Why a blended rate hides the bill
Every cascaded voice agent — Vapi, Retell, Bland, or one running on Orbit — runs the same three-stage pipeline on every turn: transcribe the caller (STT), decide the reply (LLM), speak it (TTS). A quoted "all-in" per-minute rate folds those three into one number, and the math that matters gets lost. When your invoice lands a 3–8× multiple of the advertised platform fee — the headline-vs-real gap collected in the true cost of voice AI and SMS explainer — it is because the three components passed through at three different markups, not because any of them was sneaky on its own. Component math is how you find the leak before you sign.
2. The time budget: what each stage costs in milliseconds
Cost follows time in a real-time pipeline, so the latency budget comes first. Devotel Orbit publishes a per-stage latency budget for exactly this reason, and the stage accounting looks like this per turn:
- STT — transcription of the caller's audio. Typically ~100–300 ms of the turn; it is the cheapest line on the bill and the least risky choice in the stack.
- LLM — the reasoning. Roughly ~200–800 ms on a mid-tier model, several seconds on a frontier model; it dominates both the latency and (at scale) the cost.
- TTS — speech synthesis of the reply. ~100–400 ms on a standard voice, more on a premium or cloned one; the price swings 10–30× with voice tier.
- The caller's perceived "agent speed" is the sum of all three stages plus telephony transport, so a slow stage is not just a bad answer — it is talk-minutes burned while the caller waits.
The practical read: optimize the LLM first. It holds the bulk of the elapsed turn and its token prices dwarf the other two lines the moment routing and prompts scale up.
3. The cost model: price each stage separately
Run each component as its own line, using the same published ranges as the AI voice agent pricing guide:
| Component | Metered as | Directional 2026 range | The knob that moves it |
|---|---|---|---|
| STT | Per audio minute | ~$0.005–$0.02 / min | Streaming vs batch, custom vocabulary |
| LLM | Per 1M tokens in + out | The variable line | Model tier, prompt length, caching |
| TTS | Per character of speech | ~$0.01–$0.30 / min | Standard vs premium/cloned voice |
On a typical support or booking configuration the three sum to roughly $0.05–$0.30 per talk-minute before telephony and any platform fee. The buyer's model is a four-box grid —
> LLM tier × TTS voice tier → latency profile → effective $/min
— and the point of it is disciplined substitution: swap in your vendor's actual quote cell by cell instead of trusting a blended rate, then compare honestly the way the Vapi vs Retell comparison structures its head-to-heads.
4. Worked example: mid-tier configuration
Take a 4-minute support call that runs 12 turns, with a mid-tier LLM, a standard TTS voice, and streaming STT:
- STT — 4 audio minutes × ~$0.005–$0.02 = $0.02–$0.08.
- LLM — ~600 tokens in plus ~150 tokens out per turn (~750 total): 12 × 750 ≈ 9,000 tokens. At $1–$15 per 1M tokens (mid-tier to frontier), that is $0.009–$0.135.
- TTS — ~100 characters per turn × 12 ≈ 1,200 characters; at ~1,000–2,000 characters per spoken minute that is ~0.8–1.2 speech-minutes → $0.012–$0.36 depending on the voice tier.
Sum the components and this call lands at $0.04–$0.58 per call, i.e. ~$0.01–$0.14 per talk-minute — the component way to reach the same $0.05–$0.30 all-in band published pricing guides quote. When your audit of a provider bill lands outside your own recomputed band by that much, you know which of the component lines went wrong before you spend a call with the vendor.
5. Optimization levers by stage
Each stage has one lever that dominates its bill:
- STT — stream rather than batch, and keep your custom vocabulary lean. STT is the least of the three problems.
- LLM — tier models (a small model routes and answers simple turns; a larger one fires only when the conversation earns it), cache the system prompt, and cap tool-call payloads. These three together routinely halve the token line.
- TTS — run the standard voice for routine turns and reserve the premium or cloned voice for flows where the brand register matters. Voice tier is a per-flow decision, not a per-account one.
Taken together, the levers pull a typical configuration from the top of the band toward the bottom — a 30–60% reduction in effective $/min before any vendor renegotiation.
6. Build the component ledger, then renegotiate
The AI voice agent pricing overview breaks Orbit's pay-as-you-go rate into components on purpose: a vendor that publishes only one blended number leaves you without a ledger. Before you renew or scale a voice agent contract, run this sequence:
- Ask each candidate for STT, LLM, and TTS rates as separate lines. A refusal is a signal.
- Recompute your minute using this post's formulas with your agent's average turns and prompt sizes.
- Compare the recomputed numbers against the vendor's blended quote — the gap is negotiation room, and it can be a genuine multiple.
- On Devotel Orbit, the published per-stage latency budget and component rate card make step 1 unnecessary — the four-component split is already in the open.
Frequently asked questions
Is a cascaded STT→LLM→TTS pipeline the only architecture to price?
No — speech-to-speech models collapse the pipeline into one component, and the math above still applies as a comparison baseline. Today the cascaded stack usually prices cheaper per resolved minute, because you can tier each stage independently rather than paying a frontier speech-to-speech model for every routine turn. Recheck when speech-to-speech rates move.
How do I price the LLM line before my agent is in production?
Estimate turns per call and tokens per turn from your pilot (or assume ~12 turns and ~750 tokens as above), then multiply by the published per-1M-token price of the model you will run. Re-run with production telemetry after the first week — prompts drift, and the LLM line drifts with them.
Does telephony belong in this model?
Yes, as an independent fourth line. The three compute components bill their own way, and the telephony minute (~$0.004–$0.02 on US destinations) rides under them; confusing the two is what hides the blended rate. On Orbit both lines appear on the same pay-as-you-go bill.