What a Voice AI Agent Actually Costs to Run Per Month (2026 Line-Item Breakdown)

Every voice AI vendor's homepage says roughly the same thing: "as low as $0.05 per minute." It's technically true and almost entirely useless, because $0.05/min is one line item out of five, quoted by the one vendor whose job is to make their own slice look cheap.

We build and run voice AI agents for clients, which means we also pay these bills every month: telephony invoices, LLM token usage, speech API metering, and the internal engineering time nobody puts on a pricing page. This is the breakdown we wish existed when we started pricing our first voice agent project, with every component sourced from current provider pricing pages as of August 2026, run through the math for a real call, and rolled up into what a business actually pays per month, including the part vendors leave out: the fee for the humans who keep the thing working.

The Four Things You're Paying For Every Time the Phone Rings

A voice AI agent isn't one product. It's a pipeline of separately metered services stitched together, and every one of them bills you independently the moment a call connects:

1. Telephony: carries the actual phone call over PSTN/SIP (Twilio, Telnyx, Plivo). Billed per minute, plus a monthly fee to lease the phone number.
2. Speech-to-text (STT): transcribes the caller's voice into text in real time, streaming word by word as they talk (Deepgram, AssemblyAI, OpenAI Whisper). Billed per minute of audio.
3. The LLM: reads the transcript, decides what to say, and generates the reply (Claude, GPT, Gemini). Billed per token, not per minute, which matters more than it sounds like it should.
4. Text-to-speech (TTS): turns the LLM's reply back into audio the caller hears (ElevenLabs, Cartesia, Deepgram Aura). Billed per character generated.

Some builds add a fifth layer on top: an orchestration platform (Vapi, Retell AI, Bland AI) that wires the first four together and charges its own per-minute fee for the privilege. Build it yourself and you skip that fee, but you take on the engineering work it was hiding.

Quick answer: Raw AI processing (telephony + STT + LLM + TTS) costs roughly $0.01–$0.03 per minute on a budget stack and $0.15–$0.30+ per minute on a premium one. A typical small-business voice agent handling 1,000 minutes a month therefore spends $15–$300 in raw usage, depending entirely on which models you pick. Add a platform fee (Vapi, Retell, Bland) and that climbs to $0.08–$0.40/min all-in. Add a real building and maintenance fee (the part that actually keeps the agent working) and the total client bill for a properly run 1,000-minute/month agent typically lands at $400–$2,000+ per month, of which raw compute is usually only 10–25%.

That gap between "$0.05/min" and the real invoice is the entire subject of this article. Let's go line by line.

Telephony: The Cost of Just Picking Up the Phone

Before any AI runs, someone has to carry the call. This is the least glamorous line item and, for most builds, the cheapest one.

Sourced directly from each provider's public pricing page, August 2026.

Twilio is the default most teams start with because its documentation and reliability are excellent, but Telnyx and Plivo run 30–70% cheaper for the same PSTN call because they own more of the underlying network rather than reselling capacity. For a voice agent doing meaningful volume, that difference alone is worth a proof-of-concept before you commit.

Two add-ons catch people off guard: Twilio Media Streams (the real-time audio feed your STT engine actually listens to) adds roughly $0.004/min on top of the base voice rate, and call recording + transcription can add another $0.0025–$0.05/min if you turn it on for QA or compliance. None of this is hidden. It's just easy to miss when you're only looking at the headline per-minute number.

Speech-to-Text: Turning the Caller's Voice Into Words

STT is the smallest line item in the entire stack, and it's not close.

Fetched directly from provider pricing pages, August 2026.

For a 4-minute call, STT costs somewhere between $0.01 and $0.03 total, a rounding error next to everything else in the pipeline. The real decision here isn't cost, it's accuracy versus latency. On the independent Pipecat open STT benchmark, AssemblyAI's Universal-3.5 Pro model posts the lowest word-error rate (6.99%) and is noticeably stronger at getting phone numbers and confirmation codes right: the exact entities a voice agent has to nail. Deepgram's Flux model wins on raw turn-taking latency (sub-300ms end-of-turn detection), which matters more for how natural the conversation feels than for the bill.

The LLM: What the "Brain" Actually Costs, and the One Decision That Changes It 10x

This is where most cost breakdowns get lazy and quote a single number. The honest answer is that the LLM leg can be the cheapest line item in your entire stack, or the most expensive, depending on one architectural choice made before a single call is placed.

Cascaded architecture (the standard approach: STT transcribes → LLM reads text → TTS speaks the reply) bills the LLM in ordinary text tokens. Because a voice agent's system prompt and full conversation history get resent on every single turn (LLMs are stateless), the token count per minute is higher than the words spoken would suggest. Here's the actual math, using industry-standard voice-agent assumptions (500–2,000 token system prompt, short spoken turns, ~4 turns/minute, history resent each turn):

Claude Haiku 4.5 ($1.00/1M input tokens, $5.00/1M output): at roughly 1,500 input + 250 output tokens per conversation-minute, that's (1,500 × $1 + 250 × $5) / 1,000,000 = $0.00275/min, about a quarter of a cent per minute.

GPT-5-mini ($0.25/1M input, $2.00/1M output): the same 1,500/250 token split works out to (1,500 × $0.25 + 250 × $2) / 1,000,000 = $0.00088/min, under a tenth of a cent per minute.

Now the plot twist. Speech-to-speech architecture (where a single model like OpenAI's Realtime API, gpt-realtime, listens to and speaks audio directly, skipping separate STT/TTS steps for lower latency and more natural interruption handling) prices its audio tokens completely differently: $32 per million audio-input tokens and $64 per million audio-output tokens, roughly 30–100x the per-token rate of the same provider's cheap text tier. In practice, that pushes the LLM leg of a Realtime-style agent from a fraction of a cent per minute into the $0.05–$0.20+/min range.

Neither architecture is "wrong." Cascaded is dramatically cheaper and easier to swap components in and out of. Speech-to-speech is faster to respond and handles interruptions (a caller talking over the agent) far more gracefully, because there's no transcription round-trip in the way. But if a proposal quotes you "LLM cost: $0.003/min" and then builds on OpenAI's Realtime API, the actual invoice is going to disagree with the quote by an order of magnitude, so ask which architecture you're actually paying for before you sign anything.

One honesty note: this market re-prices itself constantly. OpenAI moved through three pricing tiers on their cheap model line within about a year. Treat every number in this article as directional and check the live pricing page before you build a client quote around it. We do the same before every proposal we send.

Text-to-Speech: Usually the Biggest AI Line Item in the Whole Stack

TTS is the one most people underestimate, because "$0.05 per 1,000 characters" doesn't sound like much until you translate it into what a voice agent actually says out loud.

Here's the arithmetic: a natural speaking pace is roughly 150 words per minute, at about 5 characters per word, call it 750 characters/minute of continuous speech. In a real conversation, the AI is talking for maybe 40–50% of total call time (the rest is the caller talking, or silence). That's roughly 340 characters of generated speech per minute of call time, the number every TTS bill in this stack is actually built on.

ElevenLabs, Cartesia, and Deepgram figures pulled from official pricing pages, August 2026; per-minute figures are calculated from the character rate except where a vendor quotes a flat per-minute agent rate directly.

Two things are worth flagging. First, Cartesia's own published rate for voice-agent calling is a flat $0.06/min, notably higher than the raw character math on their per-character tiers would suggest. That premium buys their latency-optimized delivery infrastructure, not just synthesis, and it's a fair trade if sub-100ms responsiveness is the point of your build. Second, industry cost breakdowns (Inworld AI, among others) consistently find that TTS accounts for roughly 70% of the raw AI processing cost in a well-optimized voice agent stack, more than telephony, STT, and a cascaded LLM combined. If you're hunting for savings, this is where to look first, not the LLM.

Add an Orchestration Platform, and the Bill Changes Shape

Most teams don't wire Twilio, Deepgram, Claude, and ElevenLabs together by hand. They use a platform like Vapi, Retell AI, or Bland AI that bundles the pipeline and adds its own margin. Here's what each actually charges, straight from their pricing pages:

The honest read: "starting at $0.05/min" (Vapi's headline) and "$0.16–$0.40/min all-in once you add real components" (what independent comparisons consistently find once STT, LLM, and TTS are actually stacked on top) are both true statements about the same platform. The platform fee is rarely the number that surprises people. The component costs riding on top of it are.

The Real Monthly Bill: Three Worked Scenarios at 1,000 Minutes

Small-business voice agents (a receptionist or appointment-booking bot) commonly run 200–1,000 minutes a month, based on published call-volume tiers from voice-agent receptionist products. Here's what 1,000 minutes actually costs across three realistic stacks:

That first row is the number that fuels every "AI agents cost pennies" headline, and it's real, if you're an engineering team willing to self-host the orchestration, tune the prompts, monitor for silent failures, and be on call when a provider has an outage. Most businesses aren't, and don't want to be. Which brings us to the line item every pricing page leaves out.

The Part Nobody Puts on the Pricing Page: Your Building Company's Fee

Here is the uncomfortable arithmetic at the center of this whole article: raw AI usage for a 1,000-minute/month agent is somewhere between $15 and $400, depending on your stack. But real client invoices for a properly maintained voice agent typically run $400 to $2,000+ per month. That's not padding. It's the price of the part that actually makes a voice agent reliable, and it shows up in three common pricing models:

• Setup/build fee (one-time): roughly $500–$3,000 for a straightforward SMB agent (single use case, one phone line, standard integrations), rising to $2,000–$25,000+ for a complex build with CRM integrations, multi-language support, or compliance requirements.
• Monthly retainer: commonly $300–$1,500/month for an SMB agent, up to $3,500/month for more involved builds. This covers prompt iteration as call patterns reveal edge cases, monitoring for silent failures (a voice agent that mishears "cancel" as "confirm" doesn't throw an error, it just quietly does the wrong thing), and being the person who picks up when the phone system goes down at 11pm on a Friday.
• Usage markup: some shops skip the retainer and instead charge a per-minute rate that already includes their margin, commonly $0.10–$0.30/min all-in, which reads like a single clean line item on your invoice but is functionally the same fee, redistributed.

Why does this fee dwarf the compute cost? Because a voice agent isn't a script. It's a system that has to handle every way a real caller can go off the happy path: background noise, accents the STT model handles poorly, a caller who interrupts mid-sentence, a CRM API that times out, a caller who says "actually, cancel that" three sentences after confirming. None of that is solved by picking a cheaper LLM. It's solved by someone testing the agent against real call transcripts, watching it in production, and fixing what breaks. This is exactly the kind of ongoing engineering work we've written about in the context of why AI implementations fail and what custom AI software development actually costs once you look past the demo. Voice agents fail the same way, just louder: over the phone, in front of a customer, instead of quietly in a dashboard.

Hidden Costs That Blow Up the Budget Later

These rarely show up in a first quote, and every one of them is a normal, foreseeable cost, not a hidden fee, just an easily-missed one.

Concurrency/channels. Handling multiple calls at once requires provisioned capacity, not just a higher usage bill. Telnyx charges roughly $8–$12 per concurrent channel per month; Vapi's entry plan caps you at 10 simultaneous calls; Retell's free tier allows 20. Run a marketing campaign that drives a burst of calls past your concurrency limit, and new callers get a busy signal, not your AI agent.

Compliance add-ons. HIPAA-eligible configurations and zero-data-retention modes are priced separately. Vapi lists HIPAA at $2,000/month and Zero Data Retention at $1,000/month on top of usage. If you're in healthcare, legal, or finance, budget for this before you scope the project, not after a compliance review flags it.

Recording and transcription storage. Call recording for QA typically adds $0.0025–$0.05/min in processing, plus ongoing storage costs once you're past a provider's free tier.

Number porting and provisioning. Moving an existing business number onto a new telephony provider, or provisioning toll-free numbers, carries its own one-time fees and lead time, so plan for this in your launch timeline, not the week before go-live.

Fallback and monitoring engineering. What happens when the STT provider has an outage mid-call, or the LLM API times out? A production-grade agent needs fallback logic and alerting for exactly these moments. This is engineering time, not a line item on a vendor invoice, and it's one of the most common gaps in a rushed build.

How to Reduce Voice AI Costs Without Cutting Reliability

The goal isn't the cheapest possible stack. It's spending on the things that actually change the outcome.

• Cut TTS spend first, not LLM spend. TTS is roughly 70% of your raw AI cost; the LLM is often under 5%. Test a mid-tier voice (Deepgram Aura, ElevenLabs Flash) before you assume you need the most expensive multilingual model.
• Cascade instead of speech-to-speech, unless latency is the product. A cascaded Claude Haiku or GPT-5-mini stack costs a fraction of a cent per minute in LLM spend. Reach for OpenAI's Realtime API only when natural interruption-handling is a genuine requirement, not a nice-to-have.
• Use prompt caching where your context qualifies. Anthropic and OpenAI both discount repeated prompt prefixes by 75–90%. It only pays off once your system prompt clears the provider's minimum cacheable length (4,096 tokens for Claude Haiku 4.5), so this is a real design lever, not a free win by default.
• Match telephony provider to volume, not habit. Twilio's documentation is excellent for a first build; once you're past a few thousand minutes a month, a 30–70% cheaper PSTN rate from Telnyx or Plivo is worth the migration.
• Budget the retainer, don't skip it. The teams who get burned are the ones who pay for the build and assume it runs itself. It doesn't. The monthly fee is what buys you someone actually watching the failure modes above.

At Eunix Tech, we design and run voice AI agents the same way we'd want one built for us: with the real per-minute math shown up front, not a headline rate that quietly triples once every component is added. If you want an honest estimate for your call volume and use case, not a sales number, talk to us. We'll tell you what your voice agent will really cost, this month and at the volume you're planning to hit next year.

Conclusion

There's no single "what does a voice AI agent cost" number, but there is a clear way to read any quote you're given:

• Raw AI processing is genuinely cheap. It runs $0.01–$0.03/min on a budget stack, rising to $0.15–$0.30+/min on a premium one, mostly driven by which TTS voice and which LLM architecture (cascaded vs. speech-to-speech) you choose.
• TTS, not the LLM, is usually your biggest AI cost. If you're optimizing, start there.
• The architecture decision (cascaded versus OpenAI's Realtime speech-to-speech) can move your LLM cost by 10–100x. Know which one a quote assumes.
• A platform fee (Vapi, Retell, Bland) buys you speed to launch, not a lower bill. Real all-in rates commonly land at $0.08–$0.40/min once every component is stacked on top.
• The building company's fee is usually the majority of a real client invoice, and that's not padding. It's the ongoing engineering that keeps a voice agent from quietly failing in ways a dashboard won't show you.

Ask any vendor to itemize these five pieces before you sign. If they can't, that's the real answer to what it costs.

Frequently Asked Questions (FAQ)

How much does a voice AI agent cost per minute?

Raw AI processing (telephony + speech-to-text + LLM + text-to-speech) typically costs $0.01–$0.03 per minute on a budget, self-built stack, and $0.15–$0.30+ per minute on a premium stack using top-tier voices and OpenAI's Realtime speech-to-speech API. Add an orchestration platform like Vapi, Retell AI, or Bland AI, and the realistic all-in rate, after every component is stacked on top of the platform's own fee, commonly lands between $0.08 and $0.40 per minute.

What is the biggest cost driver in a voice AI agent?

Text-to-speech, not the LLM. Industry cost breakdowns consistently find TTS accounts for roughly 70% of the raw AI processing cost in a well-built voice agent, because the AI is generating hundreds of characters of speech every minute it talks, while a cascaded LLM architecture typically costs a fraction of a cent per minute. The exception is speech-to-speech architectures (like OpenAI's Realtime API), where audio-token pricing can make the LLM leg the largest cost instead.

Why is my voice AI vendor's "$0.05 per minute" quote so much lower than my actual bill?

Because that number usually covers only the orchestration platform's own fee, not the speech-to-text, LLM, and text-to-speech components running on top of it, and definitely not the building company's setup or maintenance fee. Once every component is added, realistic all-in usage costs are typically 2–8x the headline platform rate, and a full client invoice including build and maintenance work is often 5–15x higher than raw usage alone.

How much does it cost to build and maintain a voice AI agent, not just run it?

Most businesses pay a one-time setup fee of $500–$3,000 for a standard voice agent (rising to $2,000–$25,000+ for complex, multi-integration builds), plus a monthly retainer of $300–$1,500 for prompt tuning, monitoring, and fixes as real call patterns surface edge cases. This building and maintenance fee usually makes up the majority of a real monthly invoice. Raw AI usage is typically only 10–25% of what a well-run voice agent actually costs a business.

Is it cheaper to build a voice AI agent myself instead of using a platform like Vapi or Retell?

Self-hosting your own pipeline (choosing your own telephony, STT, LLM, and TTS providers directly) is genuinely the cheapest option on paper, often $15–$30 a month in raw usage for a 1,000-minute-a-month agent. But it moves the cost from a line-item fee to engineering time: someone has to build the orchestration, handle failures, and monitor quality, which platforms like Vapi, Retell AI, and Bland AI otherwise absorb into their per-minute fee. For most businesses without an in-house engineering team, the platform fee is cheaper than hiring for that work.

Does the choice of LLM really matter for voice AI agent cost?

Less than most people assume, if you're using a standard cascaded architecture. Cheap models like GPT-5-mini or Claude Haiku 4.5 cost well under a cent per minute of conversation even with a full system prompt and growing chat history resent every turn. It matters enormously, however, if you use a speech-to-speech model like OpenAI's Realtime API instead of a cascaded pipeline: audio-token pricing there runs 30–100x higher than the same provider's cheap text-token rate, which can turn the LLM into your single largest line item.

How many minutes per month does a typical small business voice AI agent use?

Based on published call-volume tiers from voice-agent receptionist products, most small-business voice agents (appointment booking, call answering, lead qualification) handle somewhere between 200 and 1,000 minutes a month. Higher-volume use cases (outbound sales dialing or high-traffic support lines) can run well into the thousands of minutes and shift the economics toward volume-discounted enterprise pricing from telephony and platform providers.