AI voice agent cost per call is the sum of four metered layers: telephony, speech recognition, the LLM and the voice. A platform fee sits on top. This guide prices each layer, works through one 3-minute call and lists the levers that cut the bill.
Four layers set your cost per call
Your voice agent is a pipeline, and each stage bills in its own unit. The carrier (SIP trunk) charges per minute. Streaming speech-to-text (STT) charges per audio minute. The LLM charges per token, in millions. Text-to-speech (TTS) charges per character, usually per thousand. A fifth line, the orchestration or platform fee, covers the software that ties the stages together.
You can't add these up until they share a unit. Convert each one to cost per connected minute, then multiply by call length to get cost per call. The LLM is awkward, because its cost depends on the number of turns.
Buyers raise pricing early, and the first question is usually per-minute billing versus a flat platform fee. Ask every vendor which of the five lines its quote includes. A single per-minute number can hide all five or only two. Per-minute pricing can also scale in ways a flat subscription does not, so model your own volume.
Telephony costs cents per minute
Telephony is the cheapest layer per minute. The rate depends on direction, destination and number type. Carriers such as Twilio, Telnyx and Vonage publish live rate tables, so check yours before you commit. The range below, $0.007 to $0.015 per connected minute, is an illustrative assumption for the worked example. It is not a quoted market rate.
Read four lines on the carrier's page:
- Outbound rate. This is the per-minute charge for calls you place. Check whether it is for a local or a toll-free destination.
- Inbound rate. This is what you pay when callers ring your number. It usually differs from outbound, and toll-free inbound is often dearer.
- Number rental. A local number carries a small monthly fee. Spread across your monthly call volume, it nearly vanishes at scale.
- Billing increment. Per-second billing charges for what you used. Per-minute billing rounds every call up, which hurts short calls most.
Rates differ across the US, the UK, the EU, India, Brazil and West Africa, and mobile numbers often cost more than landlines. Model each region you dial as its own row.
Unanswered calls are not free. Voicemail pickups bill as connected time, and some carriers also bill ringing time. Local numbers also lift answer rates, so they can pay for themselves.
Speech recognition and LLM add middle cost
STT bills per audio minute, so a 3-minute call is 3 billed minutes. Most vendors bill the open stream, not only the seconds the caller speaks. We assume the caller speaks half the call. If you gate the stream to carry only the caller's audio, you pay for 1.5 minutes. Read the streaming row, not the cheaper pre-recorded one. Deepgram's pricing page lists Nova-3 monolingual streaming at $0.0077 per minute on pay as you go. Our planning range is $0.0065 to $0.0077 per minute, where the lower figure is the page's other listed streaming rate.
The LLM bills per million tokens, with input and output priced separately. We assume speech runs about 150 words a minute and a word is about 1.3 tokens. Here is the token maths for one call:
- The call has 6 exchanges, each a caller turn of 50 tokens and an agent reply of 50 tokens.
- The system prompt is 800 tokens.
- The model re-reads the prompt and the whole history on every turn.
Input on turn k is 800 + 100 × (k − 1) + 50 tokens. Summed over six turns, that is 6,600 input tokens. Output is 300 tokens. The prompt is sent six times, 4,800 tokens, or about 73% of the input.
Prompt caching discounts that repeated prefix. Check the cached-input row on the OpenAI and Anthropic pricing pages. Cut the prompt before you chase a cheaper model.
Replace our illustrative prices with live rows: $0.15 input and $0.60 output per million tokens for a small model, and $2.50 and $10 for a large one. The call costs about $0.001 on the small model and about $0.02 on the large one. A large model is also slower, and a pause after the caller stops talking makes the agent sound broken. Route routine turns to the small model, and keep the large one for hard questions.
Voice is often the biggest metered layer
TTS bills per character the agent speaks, so price it from talk time. We assume the agent speaks half the call. At 150 words a minute and about 6 characters a word, that is roughly 900 characters per spoken minute. Half of a 3-minute call is 1.5 spoken minutes, or about 1,350 characters. Swap in your own figures.
- Premium voice: the ElevenLabs API pricing page lists $0.08 per 1,000 characters for v4, v3 and v2 Multilingual. That is about 11 cents for this call.
- Real-time voice: the same page lists $0.04 per 1,000 characters for v4 Turbo, v3 Conversational and Flash / Turbo. That is about 5 cents.
- Budget voice: Deepgram's pricing page lists Aura-2 at $0.030 per 1,000 characters on pay as you go. That is about 4 cents.
Premium costs two to nearly three times budget. Play both to real listeners before you pay the gap.
Insight7's Voice AI is in beta, so treat this as early. Insight7 Voice is a lower-cost voice engine for campaigns. It uses the same ElevenLabs voices and runs on Insight7's own infrastructure. Campaign calls can go out on local numbers, for example in Nigeria through a gateway to local carriers, because local numbers are more likely to be answered. Calls where only the AI spoke are not counted as reached. You can test an agent in the browser with no phone number.
Worked example: one 3-minute call
Price the call one layer at a time, then add them. The TTS and STT rows use the linked prices. The telephony, LLM and platform rows are illustrative assumptions, with the platform fee at $0.02 to $0.05 per connected minute. Replace them with your live rates.
| Layer | Low case | High case |
|---|---|---|
| Telephony | 3 × $0.007 = $0.021 | 3 × $0.015 = $0.045 |
| Speech-to-text | 1.5 gated min × $0.0065 = $0.010 | 3 min × $0.0077 = $0.023 |
| LLM | Small model, about $0.001 | Large model, about $0.020 |
| TTS | Budget voice, about $0.041 | Premium voice, about $0.108 |
| Platform fee | 3 × $0.02 = $0.060 | 3 × $0.05 = $0.150 |
| Total | about $0.13 | about $0.35 |
Without the platform fee, the totals are about $0.07 and $0.20. Among the metered layers, voice is the largest in both cases. The platform fee is larger still when a vendor charges one. Including the platform fee, the low case is about 4 cents a connected minute and the high case about 12 cents.
Formula: cost per call = minutes × (telephony + STT + TTS share + platform fee) + LLM token cost.
Dials that never connect still cost something. Track cost per outcome: total spend divided by successful outcomes. If one in five connected calls ends in a payment promise, each promise costs five times the call.
Six levers cut the cost per call
Start with the layer that is biggest in your own table. Seven levers do most of the work:
- Shorten the prompt and cache it. The prompt is resent every turn, so this trims LLM input.
- Use a small model for routine turns. Cost and latency both fall.
- Pick a standard voice. TTS has the widest price gap.
- Cap call length and trim the greeting. Every layer bills by the minute.
- Set calling hours and retry rules. Better answer rates mean fewer wasted dials.
- Hang up early on voicemail. Telephony and STT stop billing.
- Gate the STT stream to the caller's audio.
Don't cut blindly. A robotic voice loses callers, and buyers fear that sound. Test-call the cheap stack before launch.
Pilot on a small sample and pay on measured results, as banks ask. One ed-tech buyer wanted pricing details before a pilot, so they would not have to switch tools after committing. If you need voice fast, buy. Buying adds a platform margin. Building adds engineering cost.
Then measure what each outcome costs. Gartner found only a quarter of AI customer service use cases produce a return, so cost per outcome tells you more than cost per call. For the managed route, see Insight7's AI voice and chat agents.
Related reading: 10 Best Speech-to-Text APIs for AI Voice Agents Handling Accents and Noisy Phone Lines, Build vs Buy AI Voice Agents Cost Comparison With a 12-Month Total Cost of Ownership Model and 30 Questions a CTO Should Ask Before Building or Buying an AI Voice Agent Platform.
Questions people also ask
How much does an AI voice agent cost per minute?
Add the four layers per connected minute, then any platform fee. In our worked example the total is about 4 to 12 cents a minute. With half the call spoken, voice alone is about 1.4 cents per connected minute on Deepgram Aura-2 or 3.6 cents on ElevenLabs premium.
How much does an AI voice agent cost per call?
Multiply connected minutes by the per-minute layers, then add LLM tokens. Our 3-minute example comes to about $0.13 on a lean stack and $0.35 on a premium one. Count failed dials separately.
Which layer of an AI voice agent costs the most?
Voice synthesis leads among the metered layers with a premium voice. A large model on a long call can overtake it, because the history is resent every turn. A platform fee can top both.
Is it cheaper to build or buy an AI voice agent?
Building avoids the platform margin but adds engineering, monitoring and carrier failover work. Buying is cheaper when speed matters. Companies with their own apps and chatbots still say voice is too hard to build in-house.
Why does AI voice agent pricing differ between vendors?
Vendors bundle different layers into the quote. Some charge a flat per-minute rate, others add a platform fee, and committed volume often lowers rates.
Do unanswered calls and voicemail still cost money?
Yes, voicemail bills as connected time, and ringing can bill too. Ask your carrier about ring time, and cut voicemail calls short with early hang-up rules.
How can I reduce AI voice agent costs without hurting quality?
Change one lever at a time and compare the result. Test the cheaper stack yourself, then run a small pilot. Measure completed outcomes, not only the minute rate.
Divide spend by outcomes
Price every call as four layers plus a platform fee, then divide total spend by the outcomes you get. Judge every vendor on cost per completed outcome, because cheap calls that never convert are the dearest.

