Voice AI cost calculator

How this is calculated

Four layers, each billed in a different unit. Speech-to-text and telephony are already per-minute. The language model is priced per million tokens and text-to-speech per thousand characters, so both need a conversion from minutes — and those two conversions are where most voice cost estimates go wrong.

Two of the four layers are already priced per minute. The other two are not, and converting them is where voice cost estimates usually go wrong:

spokenWords/min = wordsPerMinute × agentTalkShare
chars/min       = spokenWords/min × charactersPerWord
outTokens/min   = spokenWords/min × tokensPerWord

sttCost         = sttRatePerMinute                       (full connected minute)
llmInputCost    = inputTokensPerMinute × rateIn  / 1e6
llmOutputCost   = outTokens/min        × rateOut / 1e6
ttsCost         = (chars/min / 1000)   × ratePer1kChars
telephonyCost   = telephonyRatePerMinute                 (full connected minute)

total/min       = stt + llmInput + llmOutput + tts + telephony + platform

Speech-to-text and telephony bill for the whole connected minute — the agent pays to listen as well as to speak. Text-to-speech and generated tokens only accrue while the agent is talking, which at a 50% talk share is about 75 words a minute.

Worked example

A lean stack — Deepgram Nova-3 for transcription, Claude Haiku 4.5, Aura-1 for the voice, inbound Twilio — comes to about $0.024 per connected minute: roughly 7 cents for a three-minute call, or $240 a month at 10,000 minutes. The split is telephony 35%, text-to-speech 26%, speech-to-text 20%, prompt tokens 17%, and generated tokens just 2%.

Swap in a frontier model and a premium voice and the same call runs about $0.062 per minute — and the shape inverts. Prompt tokens become the largest line at 32%, text-to-speech 33%, telephony drops to 23%. Which layer is worth optimising depends entirely on the stack you actually run, which is why this tool shows the split rather than just a total.

Why prompt tokens dominate and generated tokens do not

The counterintuitive result in every configuration above: the tokens the agent generates are almost free, while the tokens it reads cost real money. A minute of agent speech is roughly 100 output tokens. The system prompt, tool definitions and the entire conversation so far are re-sent on every turn, which is thousands of input tokens per minute. That ratio is why prompt caching moves the bill far more than switching to a cheaper model — and why a long system prompt is a recurring cost, not a one-off.

What this does not include

Phone number rental, recording and storage, observability, and the engineering time to assemble and keep a four-vendor stack running. A managed platform folds those into one per-minute rate — enter it as a platform fee to compare like with like. Rates here are published pay-as-you-go pricing; volume and committed-use agreements move them, generally downward.

Prices are shown in US dollars only, because every provider modelled here publishes in dollars. Converting would mean inventing an exchange rate and presenting it as a cost.

Frequently asked questions

How much does a voice AI agent cost per minute?

At published rates, a lean stack — Deepgram Nova-3, Claude Haiku 4.5, Aura-1 and inbound Twilio — comes to roughly $0.024 per connected minute, or about 7 cents for a three-minute call. A premium stack using a frontier model and a high-end voice runs closer to $0.062 per minute. Telephony and text-to-speech usually cost more than the language model on a lean stack, which surprises most people.

Which layer costs the most?

It depends on the stack, and that is the point of the breakdown. On a cheap configuration, telephony is often the single largest line — around 35% — while the language model's generated tokens are barely 2%. Move to a frontier model and a premium voice and the balance flips: the LLM's prompt tokens and TTS each take about a third. Optimising the layer that is already 2% of your bill is wasted effort.

Why are prompt tokens so much more expensive than generated tokens?

Because the system prompt, tool definitions and the whole conversation so far are re-sent on every single turn, while the agent only generates the words it actually speaks. A minute of speech is about 75 agent-spoken words — roughly 100 output tokens — against thousands of input tokens per minute. Prompt caching is the lever that matters here, not picking a cheaper model.

What assumptions does this make?

That speech runs at about 150 words per minute, the agent speaks for half the call, a word is roughly 5.5 characters and 1.35 tokens, and around 4,000 prompt tokens are sent per minute of conversation. All five are editable under "conversion assumptions" — the prompt-token figure is the one most worth replacing with a measurement from your own traffic.

Why are the prices in dollars only?

Because every provider modelled here publishes in US dollars. Converting to another currency would mean inventing an exchange rate and presenting it as a cost, which is worse than showing dollars and letting you convert with a rate you trust.

Are these the rates I will actually pay?

They are the published pay-as-you-go rates, verified on the date shown beside each one. Volume commitments, enterprise agreements and annual plans all move them, usually downward. Rates marked second-hand were reached through third-party reporting rather than the vendor's own pricing page — confirm those before making a purchasing decision.

Does this include the platform running the agent?

Only if you add one. The four component layers are what an assembled stack costs in raw infrastructure. Managed platforms bundle orchestration, telephony, monitoring and support into a single per-minute rate — enter that as a platform fee to compare like with like, and remember that a self-assembled stack also carries engineering time this model does not price.

Related reading