The true all-in cost of voice AI and SMS (not the headline rate)

A voice AI vendor's headline rate is almost always one line item in a five-line bill. Vapi's self-serve plan lists a $0.05-per-minute platform fee, but multiple independent cost breakdowns put the real all-in cost — once you add the LLM, speech-to-text, text-to-speech, and telephony that Vapi charges on top — at roughly $0.13 to $0.33 per minute, 3–8× the advertised number. This is not unique to one vendor: it's structural to how "bring your own AI stack" voice platforms price, and it shows up in SMS pricing too, just through a different mechanism (registration fees, carrier surcharges, and retried failed sends). This guide breaks down where the gap comes from and gives you a calculator to estimate your own real number.

Where the headline-vs-real gap comes from in voice AI

A voice AI call is not one billed service — it is at minimum five, often billed by four different companies:

1. Platform / orchestration fee. The number vendors headline — for hosting the call, managing the WebSocket connection, and stitching the other components together.
2. Speech-to-text (STT). Transcribing the caller's audio in real time, typically ~$0.01/minute.
3. The LLM. The reasoning layer, and the biggest swing factor — a frontier model can run $0.08–$0.20/minute depending on prompt length and context, while a smaller model can bring it under $0.05/minute.
4. Text-to-speech (TTS). Synthesizing the agent's voice, typically $0.02–$0.05/minute depending on voice quality.
5. Telephony / termination. The actual PSTN minute — carried separately from the AI stack.

Independent cost breakdowns of Vapi's stack converge on the same range from different angles: one analysis based on production deployments puts a budget stack (smaller LLM + fast STT/TTS) around $0.15/minute and a premium stack (frontier LLM + higher-fidelity voices) at $0.35–0.40/minute; another, based on 40+ deployments, lands on $0.15–$0.31/minute before telephony; a third puts the realistic range at $0.15–$0.36/minute, commonly $0.23–$0.33. None of this is a vendor being dishonest about its own line item — $0.05/minute is a true number for what it covers. The gap exists because most public "voice AI pricing" comparisons quote only the platform fee, not the full stack a production call actually runs on.

Where the gap comes from in SMS

SMS headline rates hide cost differently — not through unbundled components, but through fees that don't show up on the per-message rate card:

• Carrier registration fees. US A2P 10DLC requires brand and campaign registration with per-carrier fees that aren't part of the advertised per-segment price.
• Failed and retried sends. A message that fails delivery and is retried is often billed again, which a simple "price × volume" estimate misses entirely.
• Artificially inflated traffic (AIT). A portion of international A2P traffic is fraudulent, and brands pay for it at the published rate regardless — see our companion guide on AIT and SMS-pumping costs.
• Currency and destination surcharges. A single "$X per SMS" headline rate typically applies to one reference country; costs to harder-to-reach destinations are frequently several times higher.

Estimate your own real all-in cost

The calculator below lets you enter your own monthly call volume and adjust each cost component to see the gap between a headline platform fee and the real all-in cost, using the ranges independent testing has published for a typical voice AI stack.

How to read a vendor's rate card honestly

1. Ask what the quoted number excludes, not just what it includes. "Platform fee" and "all-in cost" are different questions.
2. Price your own stack, not a vendor's demo stack. LLM cost depends heavily on your prompt length and model choice — a longer system prompt or a frontier model can double the reasoning-layer cost alone.
3. Model your failure/retry rate for SMS, not just the happy-path send cost.
4. Ask for a single combined rate when it exists. A platform that bills the full stack as one line (rather than passing through four separate vendor invoices) removes the reconciliation problem, even if the total number is similar.

Orbit's angle: one line, not five

Orbit bills AI voice agents and SMS as a single pay-as-you-go rate per channel — the LLM, STT, TTS, and telephony legs of a voice AI call are not separate invoices from separate vendors, and the same account carries every other channel too. Current published per-minute and per-message rates are on the pricing page, and outbound voice terminates over Devotel's own wholesale softswitch rather than a resold aggregator hop — see the true cost breakdown for AI voice agent pricing and Orbit vs Vapi for a direct comparison.

Frequently asked questions

Is a vendor lying if their headline rate doesn't include the LLM?

No — a "$0.05/minute platform fee" is typically an accurate description of what that specific line item covers. The issue is that most public pricing comparisons quote only that line item and present it as the full cost of running a voice agent, when independent testing consistently shows the full stack costs several times more.

What is the single biggest driver of the real-vs-headline gap in voice AI?

The LLM. Independent breakdowns show it can range from under $0.05/minute (smaller models) to $0.20/minute (frontier models with long prompts) — a 4× swing on one component alone, before STT, TTS, and telephony are added.

Does this gap apply to every voice AI vendor, or just one?

The structural gap (platform fee vs. full stack cost) applies to any "bring your own AI stack" model, where you separately select and are billed for your LLM, STT, and TTS providers. It is less relevant to a platform that bills the full stack as one combined rate rather than passing through unbundled third-party invoices.

How does SMS hide cost differently from voice AI?

SMS pricing is usually quoted as a single per-segment rate, so the hidden cost isn't an unbundled component — it's fees that don't appear on the rate card at all: carrier registration, retried failed sends, and artificially inflated (fraudulent) traffic that still bills at the published rate.

How accurate is the calculator on this page?

It uses published, independently-tested ranges for each voice AI stack component (see the Sources below) as starting defaults, but every field is editable — enter your own vendor's quoted rates for an estimate specific to your stack. It is an estimation tool, not a quote.

Sources and further reading

• Vapi Pricing 2026: What It Costs & How to Think About It — Emitrr: independent breakdown of Vapi's platform fee vs. full-stack cost.
• Vapi Pricing 2026: $0.05/min Listed, $0.15-0.40 Real Cost — pxlpeak: cost ranges based on production deployment testing.
• What the Per-Minute Model Actually Costs — HappyRobot: independent testing of real-world per-minute cost.
• Vapi Pricing Demystified — Five.co: component-by-component breakdown of the five billed services in a voice AI call.

Published 26 July 2026. Part of the Orbit resources library — foundational guides for teams building on communications infrastructure.