Research-backed calculator guide

Voice Agent Cost per Minute: Build an All-In Stack Before You Compare Platforms

Use a provider-neutral voice agent cost per minute calculator method to price STT, LLM, TTS, telephony, platform fees, fixed costs, and uncertainty.

What a voice agent cost per minute calculator must include

1. Speech-to-text: price the audio that is actually processed

STT usually follows caller audio minutes, not the number of words in the transcript. If the recognizer listens during silence, hold time, or barge-in, those seconds may still count under that provider’s rules. If audio has multiple channels, check whether the provider bills total processed channel duration rather than wall-clock duration. Deepgram states that streaming STT is billed per minute, with Nova-3 monolingual at $0.0077 per minute on pay-as-you-go, and says its normal audio requests are billed per second without block rounding ^[1].

For a calculator, use call minutes × STT rate and add optional features separately. Diarization, redaction, keyterm prompting, or a premium multilingual model can change the rate. Do not silently substitute a batch transcription price for a real-time call price.

2. LLM: count input replay as well as visible output

The LLM line is often underestimated because the visible answer is only part of what the model reads. A live agent may resend a system prompt, tool definitions, conversation history, the new caller turn, and metadata on every turn. Input tokens can therefore rise faster than call length when the full history is replayed. Output tokens normally track how much the agent speaks.

Use separate fields for input and output. OpenAI’s pricing page shows standard models priced per 1 million tokens and lists audio models separately; it also states that backend model and tool usage is charged separately for per-second voice sessions ^[2]. That means a “voice model” label does not automatically mean every tool call, prompt token, or external service is included.

A simple worksheet can approximate input tokens as:

input tokens = repeated prompt tokens + cumulative transcript tokens seen across turns

For higher confidence, export token usage from a test call. Use the calculator for a planning estimate, then reconcile the estimate against real provider usage.

3. Text-to-speech: convert the agent’s words into characters

TTS is commonly billed on input characters, while people plan in minutes of conversation. Those are not interchangeable. A call in which the caller does most of the talking can use much less TTS than one in which the agent narrates every step. Define the agent speech share and characters per word before converting.

TTS characters = agent words × characters per word

ElevenLabs lists v4 Turbo, Flash/Turbo, and v3 Conversational at $0.04 per 1,000 characters on its API pricing page; it lists other models at different rates and prices Scribe speech-to-text by hour ^[3]. Use the selected model’s actual unit, not a generic “per-minute voice” assumption. If the provider quotes a connected-minute speech product, do not add a second TTS charge for the same audio.

4. Telephony or WebRTC: separate transport from intelligence

Phone calls add a carrier meter. Twilio’s U.S. pricing page lists local outbound at $0.0140 per minute, local inbound at $0.0085, local number rental at $1.15 per month, and other optional features such as recording or Conversation Relay ^[4]. Destination, direction, SIP, number type, recording, and transfers can change the total.

Browser-based WebRTC may have a different participant-minute or bandwidth meter. Keep the transport field explicit so that a browser agent is not accidentally priced as a PSTN call. For transfers, decide whether the AI continues to run during the handoff and whether the human leg has a separate carrier charge.

5. Platform, hosting, and fixed fees

A raw component sum is not automatically an all-in number. Add the orchestration or platform fee if a managed service charges one. For self-hosting, include compute, media servers, storage, monitoring, and the concurrency needed to meet your latency target. Allocate fixed costs to usage:

fixed fee allocation per minute = monthly fixed fees ÷ expected monthly billable minutes

This allocation is sensitive to utilization. A $1.15 phone number contributes $0.000115 per minute at 10,000 monthly minutes, but $0.00115 per minute at 1,000 minutes. The fee did not change; the denominator did.

Worked cost example: a three-minute U.S. outbound call

This is an illustrative calculation, not a quote. It uses published rates where available and clearly marked planning assumptions elsewhere. The call lasts three minutes, has 150 spoken words per minute, four turns per minute, and the agent produces 45% of the spoken words. We assume 1.3 tokens per word, five characters per word, an 800-token prompt and tool package repeated on each turn, and 10,000 monthly minutes for allocating a $1.15 phone number. The LLM rates use OpenAI gpt-6-luna’s listed $0.10 per million input tokens and $0.50 per million output tokens ^[2]; the STT, TTS, and telephony rates come from the cited provider pages ^{[1] [3] [4]}. The $0.01 per minute platform allowance is a planning assumption, not a vendor rate.

The token arithmetic is intentionally visible. The 12-turn call repeats the 800-token prompt 12 times, or 9,600 tokens. The 652.5 total spoken words become 848.25 transcript tokens. Assuming each transcript token is seen an average of 6.5 times as history, the cumulative transcript portion is 5,513.6 tokens. Together they produce about 15,113.6 input tokens. This is a model of history replay, not a guarantee of any provider’s internal accounting.

Rounded to normal budgeting precision, the base result is $0.138 per call. Dividing by three minutes gives $0.046 per minute. Add a 10% planning allowance for retries, small rate changes, unmodeled add-ons, and call-shape variance: $0.13759 × 1.10 = $0.15135, or about $0.050 per minute. Keep this allowance separate from the vendor rates so finance can remove it after measuring production variance.

If the same call is inbound rather than outbound on Twilio local numbers, replace $0.04200 with $0.02550 for three minutes. If the agent speaks 70% of the time, TTS rises materially. If the call uses a bundled realtime product, remove the component meters that the bundle explicitly includes before adding the bundle rate. Never add a bundle and its included STT, TTS, or LLM again.

How to compare platforms without mixing units

Start every comparison with a normalized call card: direction, country, duration, silence policy, agent speech share, expected turns, model, voice, STT mode, transfer behavior, and whether recording or analytics are enabled. Then map each quote to one of four categories: included, separately metered, fixed monthly, or unknown. An unknown line is not zero.

Use the model pricing comparison when the main uncertainty is model choice. Use the AI API invoice reconciliation calculator after launch to compare forecast meters with actual invoice rows. If monitoring and trace retention are part of the deployment, estimate their effect with the AI observability ROI calculator. These links are useful because a per-minute forecast is only reliable when it can be reconciled to usage and operational outcomes.

Assumptions, exclusions, and limitations

This worksheet assumes English, United States pricing and a U.S. local outbound call. Taxes, regulatory surcharges, international destinations, number-porting costs, SIP setup, carrier minimums, support plans, enterprise commitments, and volume discounts are excluded unless explicitly added. Published pages can change, and some prices are plan-specific or promotional. Check the linked provider page before signing a contract.

The example also excludes human labor, prompt development, CRM fees, knowledge-base retrieval, external tools, payment processing, fraud controls, compliance reviews, and failed-call or dial-attempt charges. It treats platform cost as a $0.01-per-minute estimate solely to demonstrate how an unquoted orchestration layer belongs in the model. Replace that estimate with a vendor quote or measured infrastructure cost.

Silence and history handling are the largest behavioral uncertainties. A provider may meter connected time, processed audio, generated characters, tokens, or a bundled session. Barge-in can reduce generated speech but increase interruption events. Transfers can stop AI billing while telephony continues. Record these rules next to the rate, with an effective date, so a later comparison remains auditable.

FAQ

How much does a voice agent cost per minute?

There is no single market rate. A raw component stack can be a few cents per minute, while a managed or premium stack can be materially higher. Calculate STT, input and output LLM tokens, TTS characters, transport, platform or hosting, and fixed-fee allocation separately. The worked example here is about $0.046 per minute before a 10% planning allowance, but its result depends on the stated call shape and provider choices.

What is the biggest cost driver in a voice bot cost per minute breakdown?

It depends on the call shape. Agent-heavy conversations make TTS important because more characters are generated. Long, tool-heavy conversations can make LLM input important because history and tool definitions are replayed. PSTN calls can make telephony comparable to model spend. Run a sensitivity check for call length, agent speech share, turns per minute, and prompt size instead of assuming one line always dominates.

Should I use a bundled voice-agent price or add every API line?

Use one method at a time. If a bundle explicitly includes STT, LLM, TTS, and infrastructure, use the bundle rate and add only excluded telephony, numbers, transfers, or add-ons. If you build the stack yourself, add the component rates. Double-counting included components is one of the most common calculator errors.

Do voice-agent calculators charge for silence?

Not universally. Some meters follow connected call time; others follow processed audio or a platform-specific session rule. Read the billing definition for the selected service and model silence trimming explicitly. If the rule is unknown, show a low and high case rather than treating silence as free.

How do I convert a monthly estimate into cost per minute?

First calculate total monthly spend, including fixed fees. Then divide by billable minutes, not merely attempted calls. For example, $1,000 of fixed and usage spend across 10,000 billable minutes is $0.10 per minute. If volume falls to 1,000 minutes, the same fixed cost adds $1.00 per minute before usage.

What should I validate after launch?

Export call duration, STT seconds, input and output tokens, TTS characters, telephony minutes, transfers, retries, and platform charges. Compare actual totals with the calculator by component. Use the AI API invoice reconciliation calculator to investigate meter differences, then update the call-shape assumptions rather than applying an unexplained fudge factor.