Research-backed calculator guide

Voice AI Cost per Call: Convert Per-Minute Rates Into a Finished-Call Budget

Learn how to calculate voice AI cost per call from provider meters, transfers, fixed fees, and call outcomes, with a worked U.S. example.

What belongs in voice AI cost per call?

Start by defining the unit. An answered call is a connected call that reached the agent. An AI-resolved call ended without a human handoff. A transferred call includes the agent's work plus the downstream human conversation. A dial attempt includes unanswered, busy, failed, or voicemail outcomes, depending on your outbound workflow. Do not call these interchangeable “calls” in a budget sheet.

A useful call ledger has one row per call or one row per metered call segment. At minimum, record call ID, direction, connected status, AI seconds, transfer seconds, provider, model, voice, add-ons, outcome, and billed amount. This makes it possible to reconcile the estimate to an invoice later; the site's AI API invoice reconciliation calculator is useful when provider usage and invoice rows no longer line up.

1. AI runtime

The first line is time during which the AI stack is active. Depending on the platform, the rate may bundle infrastructure, speech recognition, a language model, and speech synthesis, or expose each component separately. Retell's pricing page lists a $0.055 per-minute voice-infrastructure charge, model-specific LLM rates, text-to-speech rates, telephony, and optional add-ons. It also advertises an overall AI voice-agent range of $0.07–$0.31 per minute. Treat that range as a starting point, not as the cost of every possible call configuration. ^[1]

Vapi describes a different meter structure: its platform hosting is $0.05 per minute, while transcriber, model, voice, and telephony provider costs are charged at cost. Its pricing calculator shows those lines separately. That means a Vapi estimate should not stop at $0.05 per minute; add the selected provider rates and any success package or transport fees. ^[2]

2. Speech and language meters

A custom stack can bill speech-to-text by audio time, text-to-speech by generated characters, and the LLM by input and output tokens. These units do not grow in the same way. A longer call increases STT minutes, but it can increase LLM input tokens faster if conversation history is replayed on every turn. Agent verbosity increases TTS characters, while a silent caller may still create listening-time cost.

Deepgram's published pricing illustrates why the unit matters: its streaming STT table quotes rates per minute, its TTS table quotes per 1,000 input characters, and its managed Voice Agent table quotes one rate per minute. The page also says normal audio is billed by the second without block rounding. Use the exact meter for the product you selected rather than copying a rate from a different API. ^[3]

ElevenLabs presents another caution. Its general plans are expressed in monthly credits and include different amounts of text-to-speech and other products. A credit balance is not automatically a live-call minute. Convert it only after you know the model, character consumption, agent speech share, and subscription allocation. The pricing page separately identifies agent/API products and plan limits, so keep subscription allocation visible in your model. ^[4]

3. Telephony and transfers

Telephony can be embedded in an all-in-one rate, charged separately by the platform, or billed by your carrier. For a U.S. baseline, Twilio's pricing page lists local outbound calls at $0.0140 per minute, local inbound calls at $0.0085 per minute, and local-number rental at $1.15 per month as of the page's August 2026 pricing note. Toll-free and feature rates differ. ^[5]

A transfer deserves its own fields. Some platforms stop the AI-agent meter when the destination answers but continue telephony for the human leg; Retell states this treatment in its pricing FAQ. A warm-transfer briefing or hold may still be part of the AI leg. Your calculator should therefore store AI duration and post-transfer duration separately instead of multiplying one average call length by one blended rate. ^[1]

4. Fixed and per-event charges

Allocate monthly charges only after deciding the denominator. A $2 phone number spread over 500 answered calls contributes $0.004 per answered call. The same number spread over 50 calls contributes $0.04. Concurrency, knowledge bases, verified numbers, recording, storage, branded calling, batch dialing, quality assurance, and CRM or contact-center features may also be fixed or per-event charges. If you cannot attribute a fee to a call, label it “allocated overhead” rather than hiding it inside the per-minute rate.

The per-call calculation method

Use this sequence for each call type:

1. Choose the outcome denominator. Decide whether the KPI is cost per answered call, cost per AI resolution, cost per transfer, or cost per connected outbound conversation.
2. Normalize meters. Convert seconds to minutes, characters to the selected TTS unit, and token charges to dollars. Preserve the original units beside the normalized value.
3. Separate AI and human legs. Record AI-active minutes and post-transfer minutes independently.
4. Add event charges. Include batch dials, branded caller ID, recording, or other fees that occur once per call.
5. Allocate fixed charges. Divide only the relevant monthly fixed costs by the chosen monthly call denominator.
6. Reconcile actuals. Compare the estimate with provider usage and invoices. Investigate duration, model, voice, transfer rate, and prompt length before changing the assumed rate.

For a monthly forecast, use:

monthly cost = (AI calls × average AI minutes × AI rate) + (transferred calls × post-transfer minutes × telephony rate) + event charges + fixed fees

Then use:

cost per answered call = monthly cost ÷ answered calls

cost per AI-resolved call = monthly cost ÷ AI-resolved calls

The second denominator can look surprisingly high because transferred calls still cost money but are excluded from the “AI-resolved” count. That is not a mathematical error; it is a stricter unit-economics measure.

Worked cost example

Assume a U.S. inbound support line with 500 connected calls per month. The agent spends 3.2 minutes on each call. Eighty percent end with the AI, while 20% transfer to a person for another 4.5 minutes. Use this illustrative stack, based on the component shape shown on Retell's pricing page: $0.055 per AI voice-infrastructure minute, $0.040 per LLM minute, $0.015 per platform-voice minute, and $0.015 per telephony minute. The resulting AI-active rate is $0.125 per minute. ^[1]

The following table makes every assumption inspectable. The figures are planning estimates and exclude human labor, implementation, taxes, carrier contracts other than the stated telephony assumption, and revenue or quality outcomes.

The result is not “the” voice agent average cost per call. It is the output of a transparent call shape. If average AI time rises from 3.2 to 4.0 minutes, the runtime line becomes $250 before transfer and fixed costs. If the transfer rate rises, the extra telephony line rises even when the AI duration does not. If your vendor rounds every call to a minute instead of accumulating seconds, short-call results will differ again.

For provider selection, place the same call shape into the model pricing comparison, then compare the resulting rate and unit definition rather than comparing headline numbers alone. If you use separate STT, TTS, and LLM services, the Mistral and Groq API cost calculator can help with model-rate scenarios, but you still need to add telephony and call-outcome fields for a complete phone budget.

How to make the estimate closer to the invoice

Use connected-call data, not a generic duration. Pull median and average AI duration from your phone system. The average captures spend; the median reveals whether a small number of long calls are distorting the plan.

Measure agent speech share. TTS may be billed by characters or audio duration. A talkative agent and a caller-led agent with the same call length can have different synthesis costs.

Track prompt and turn growth. Long prompts, tools, and repeated conversation history can increase LLM input tokens. A small model-rate change may matter less than adding one unnecessary confirmation turn to every call.

Keep transfer phases visible. Record warm-transfer briefing, hold, destination connection, and post-transfer duration as separate timestamps. Never assume a handoff makes the whole call free of AI or telephony charges.

Separate estimate, quote, and actual. A published rate is not a custom enterprise quote. A calculator output is an estimate. An invoice is the actual to reconcile. Mark all three in your ledger and date pricing observations because provider pages change.

Model failed and voicemail outcomes explicitly. Retell says failed calls that never connect are not billed and voicemail is billed only while the AI is active, but another provider or carrier may define outcomes differently. Do not generalize that rule across vendors. ^[1]

Review overhead with an outcome metric. The AI observability ROI calculator is relevant when the question shifts from “what did this call cost?” to “what level of monitoring, evaluation, or retention cost is justified by the operational result?” Keep that layer separate from the direct call meter.

FAQ

Is voice AI cost per call just the per-minute price multiplied by call length?

Only when the provider has one genuinely all-in per-minute meter and no transfer, event, or fixed charges. Most stacks have at least two units. Multiply AI-active time by the AI rate, add post-transfer telephony where applicable, add per-call charges, and allocate fixed fees to a clearly named denominator.

What is a reasonable AI phone agent cost per call?

There is no universal benchmark that applies to every stack. A three-minute call at $0.10 per AI minute costs $0.30 before fixed or transfer charges; the same duration at $0.30 costs $0.90. The useful answer is the rate and call shape you can substantiate. Record model, voice, telephony, add-ons, billing precision, and date rather than publishing one unexplained average.

Should I divide by answered calls or completed calls?

Use answered calls for the average cost to serve a connected caller. Use AI-resolved calls for the stricter cost of an outcome completed without handoff. Report both when transfers are material. Dividing by only successful AI resolutions assigns the cost of transferred calls to the resolutions that remain, which can be useful for unit economics but should be labeled clearly.

Do transferred calls cost more than AI-only calls?

Usually they can, because the AI leg remains in the cost and an additional human-leg telephony duration may be billed. The exact rule depends on the platform and carrier. Retell states that its AI fee stops after transfer while telephony continues, so a transferred call should retain AI minutes before the handoff and add the downstream telephony minutes. Verify the rule for your provider before modeling. ^[1]

How do I price a custom stack with STT, TTS, and an LLM?

Keep the meters separate. Multiply STT audio minutes by the STT rate. Multiply generated characters by the TTS rate per character or per 1,000 characters. Multiply input and output tokens by their respective rates. Add hosting, telephony, retries, storage, and fixed fees. This is more reliable than forcing a character- or token-based provider into a single guessed per-minute rate.

How often should I update a voice AI cost worksheet?

Update it when a provider price, model, voice, telephony route, prompt, transfer policy, or call mix changes. In production, compare the forecast with the first two to four weeks of usage and then recheck monthly. Date every quoted rate and distinguish a planning estimate from an invoice-backed actual.