Voice AI Agent Cost Per Minute: The Real Math

A voice AI agent does not cost $0.05 per minute. That figure, which several platforms advertise, is one line of a five-line invoice, and the other four lines are often larger. A realistic production agent in late 2026 costs between $0.06 and $0.31 per connected minute, depending on how you assemble the stack and which model you put in the middle of it.

This post breaks the per-minute price into its five billing layers, verifies current rates from the vendors' own pricing pages, runs a 10,000-minute worked example four different ways, and covers the charges that never make it onto the pricing page. If you are scoping a voice agent budget, this is the math your vendor's estimator rounds away from.

The five meters inside every voice AI minute

Every voice agent call passes through the same pipeline: a phone line, a speech-to-text engine, a language model, a text-to-speech engine, and an orchestration layer that manages turn-taking and tool calls. Each layer bills separately unless you buy a bundle.

Telephony: $0.003 to $0.022 per minute. This is the call itself. Twilio charges $0.0085 per minute for US local inbound and $0.0140 for outbound; toll-free inbound is $0.0220 (Twilio pricing). Telnyx SIP trunking starts at $0.0032 per minute inbound and $0.005 outbound (Telnyx pricing). On Vapi, transport is pass-through: Twilio inbound at $0.008 per minute, Telnyx at $0.0055, or free if you use Vapi's own telephony (Vapi pricing).

Speech-to-text: $0.0065 to $0.01 per minute. Deepgram Nova-3 streaming, the model behind most real-time agents, lists at $0.0077 per minute pay-as-you-go and $0.0065 on the Growth plan, billed per second (Deepgram pricing). ElevenLabs Scribe v2 Realtime runs $0.39 per hour, or about $0.0065 per minute (ElevenLabs pricing). Add-ons cost extra: redaction and speaker diarization add $0.002 per minute each at Deepgram.

LLM: $0.002 to $0.16 per minute. The widest range of any layer. Retell's component menu prices GPT 5 nano at $0.0016 per minute, Claude 5.5 Haiku at $0.0032, Claude 5 Sonnet at $0.064, and Claude 5 Opus at $0.16 (Retell pricing). These per-minute figures assume a realistic production prompt of around 6,000 tokens per minute including conversation history, tools, and retrieved knowledge. The model you pick moves your bill more than anything else.

Text-to-speech: $0.015 to $0.10 per spoken minute. Retell charges $0.015 per minute for platform voices (Cartesia, OpenAI, Inworld), $0.040 for ElevenLabs Flash, and $0.10 for ElevenLabs v3 or v4. ElevenLabs' own API bills per character: Flash and Turbo cost $0.04 per 1,000 characters, v4 costs $0.08. Since one spoken minute is roughly 750 to 900 characters (Bearworks cost analysis), that is $0.030 to $0.036 per minute the agent actually talks, and less per call minute if the caller does half the talking.

Orchestration: $0.05 to $0.055 per minute. The platform fee for turn-taking, interruption handling, tool calls, and streaming. Vapi charges $0.05 per minute of hosting with models passed through at cost. Retell's voice infrastructure layer is $0.055 per minute. This is the number most platforms advertise, because it is the only one they fully control.

What the major platforms charge right now

Pricing pages disagree with each other, partly because each vendor defines the bundle differently. Here is what each actually sells, from their own pages as of November 2026.

Telnyx sells a $0.05 per minute voice engine that includes orchestration, hosted STT, and hosted TTS, then adds LLM tokens on its own GPUs (it quotes about $0.006 per minute for a Kimi model) and telephony at carrier cost. Its own worked example lands at roughly $0.06 per minute all-in, $13,700 per month at 230,000 minutes. One footnote matters: Telnyx's estimates exclude per-call rounding to 60-second increments, so short calls cost more than the sticker.

Vapi charges $0.05 per minute for hosting and passes every model through at cost. Its own estimator prices Deepgram at $0.0095 to $0.0099 per minute, OpenAI models at $0.0077 to $0.0452, and ElevenLabs voices at $0.0146 to $0.0238. Platform packages sit on top: Core at $29 per month, Pro at 10 percent of your hosting fee with a $999 monthly minimum. HIPAA handling is a $2,000 per month add-on.

Retell sells all-in bundles from $0.07 to $0.31 per minute, and publishes the full component menu so you can assemble your own. Its default estimate, $0.055 infrastructure plus $0.015 platform TTS plus $0.080 LLM, lands at $0.15 per minute. Swap the LLM for GPT 5 nano and the same agent drops to about $0.072.

Deepgram offers its Voice Agent API at a flat $0.075 per minute pay-as-you-go for a standard stack, $0.163 for the advanced tier, and $0.050 if you bring your own LLM and TTS.

ElevenLabs prices its speech-to-speech Realtime v1 Mini at $0.08 per minute plus LLM usage.

The pattern to notice: a $0.05 sticker means orchestration only at Vapi, orchestration plus speech at Telnyx, and nothing at all at platforms that lead with a model price. Always ask what is inside the number before comparing numbers.

Worked example: 10,000 minutes a month, four ways

Take a realistic workload: an inbound support agent handling 10,000 connected minutes per month on US local numbers. Four ways to build it, using the verified rates above.

Option 1, bundled on Telnyx: about $596 per month. Voice engine at $0.05, LLM on Telnyx GPUs at $0.0064, SIP inbound at $0.0032. Total $0.0596 per minute. The cheapest option here, and one invoice instead of five.

Option 2, assembled on Vapi: $898 to $1,369 per month. Hosting $0.05, Deepgram $0.0095, a mid-tier OpenAI model around $0.02, ElevenLabs voice $0.015, Twilio inbound $0.008. That midpoint of $0.098 per minute assumes average model choices; a frontier model and premium voice push you toward $0.137.

Option 3, Retell all-in: $716 to $1,340 per month. With GPT 5 nano and platform voices, $0.0716 per minute. With Claude 5 Sonnet, $0.134 per minute. Same platform, a 1.9x cost swing from one dropdown.

Option 4, self-hosted: about $1,324 per month, and it is the wrong choice at this volume. A Bearworks engineering analysis of self-hosting LiveKit with commercial APIs puts variable cost at $0.0124 per minute but adds roughly $1,200 per month of fixed Kubernetes infrastructure. At 10,000 minutes, that fixed cost alone is $0.12 per minute. At one million minutes per month, the same stack costs $13,600 total, or $0.0136 per minute, against $158,629 for an equivalent managed stack. Self-hosting is a volume game; below roughly 100,000 minutes a month, the fixed costs eat the savings.

Notice what the worked example shows: at 10,000 minutes, the spread between the cheapest and most expensive sane build is about $750 per month. The model and voice choices matter more than the platform choice.

The charges that do not fit on a pricing page

The per-minute rate is the beginning of the invoice, not the end. These are the meters buyers miss, and every one of them is documented on a vendor page or FAQ.

Silence and hold time are billed. Retell states it plainly: billing covers the entire call duration because the speech-to-text engine stays active and listening even during silence. If your average call is six minutes of talk plus two minutes on hold while the agent checks a backend system, you pay for eight. That is a 33 percent premium over the talk-time estimate.

Per-call rounding punishes short calls. Telnyx's own footnote admits its estimates exclude rounding to 60-second increments. A 15-second "wrong number" call bills as a full minute on platforms that round per call. Retell and Deepgram bill per second; check before you assume.

The LLM re-sends the whole conversation every turn. As AssemblyAI's pricing analysis explains, each turn of a call re-sends the system prompt, the conversation history, retrieved knowledge, and tool definitions. Token usage grows with call length, so an eight-minute call costs more than double a four-minute call on the LLM line. Prompt caching helps, but only if your provider passes the discount through.

Tool calls add invisible tokens. Every lookup the agent performs is a round trip through the model. An agent that checks an order status, verifies an account, and books an appointment makes three tool calls per conversation, each one adding tokens to both the input and output lines.

Concurrency is a separate meter. Vapi's free tier includes 4 concurrent calls; extra lines cost $10 per line per month. Retell includes 20 free concurrent calls, then $8 per line. If your call pattern is bursty, 200 calls a day arriving in a two-hour window needs far more lines than the same 200 spread evenly.

Compliance costs extra. HIPAA-eligible handling on Vapi is a flat $2,000 per month add-on. Retell gates PII redaction, custom data retention, and BAAs behind its enterprise plan.

Small per-call fees accumulate. Twilio charges $0.0075 per call for answering machine detection and $0.0025 per minute for call recording. On 100,000 outbound dials, AMD alone is $750.

Platform minimums change the math at low volume. Vapi Pro has a $999 monthly minimum, Telnyx Growth a $2,000 minimum, Deepgram Growth a $4,000 annual prepay. At 10,000 minutes a month these floors can exceed your usage charges entirely.

Cost per resolved call: the number that actually matters

Per-minute cost is an input. The metric your CFO cares about is cost per resolved task, and it moves in the opposite direction from per-minute cost often enough to matter.

The formula is simple: monthly cost equals calls times average connected minutes times your per-minute rate, and cost per resolution divides that by the number of calls actually completed without human help. AssemblyAI's example: 30,000 calls averaging four minutes at $0.075 per minute costs $9,000 per month. If the agent resolves 24,000 of them, that is $0.375 per completed task.

Now run it the other way. A premium stack at $0.15 per minute that resolves 85 percent of calls can beat a cheap stack at $0.07 that resolves 60 percent and transfers the rest to humans, because every transfer adds the human agent's fully loaded cost per call. When you evaluate vendors, ask for resolution rate and per-minute cost together. Either number alone is marketing.

How to price your own agent in 30 minutes

You can build a defensible estimate before you talk to a single vendor.

1. Measure your real call profile. Pull three months of call data: total connected minutes, average duration, the share of calls under one minute, hold time as a percentage of call time, and peak concurrent calls. Every one of these maps to a line item above.
2. Pick your model tier honestly. Simple FAQ and booking agents run fine on the $0.002 to $0.008 per minute models. Agents that reason over documents or handle angry customers need the $0.06-plus tier. Do not benchmark the cheap model and budget for it if you will ship the expensive one.
3. Choose the voice for the use case, not the demo. Platform voices at $0.015 per minute are fine for transactional calls. Premium voices at $0.08 to $0.10 per minute are for brand-facing lines. The demo always uses the premium voice.
4. Add the hidden multipliers. Multiply your talk minutes by one plus your hold-time share. Add 20 to 40 percent to the LLM line if your agent makes tool calls or handles long conversations. Add per-call fees: AMD, recording, rounding.
5. Check the floors and caps. Compare your monthly total against platform minimums, and your peak concurrency against the included lines. A $500 usage month on a $999-minimum plan is a $999 month.
6. Run the resolution-rate sanity check. Estimate cost per resolved call for your top two vendor options, not just cost per minute. The cheaper per-minute option loses if it resolves noticeably fewer calls.

When each pricing model is the wrong choice

Bundled flat-rate is wrong when you need a specific premium voice or a frontier model the bundle does not include, or when your call volume is high enough that the bundle's convenience premium exceeds a self-managed discount. It is also wrong if the vendor rounds per call and your calls are short.

Assembled pay-per-component is wrong when your team cannot operate five vendor relationships, when you need a single throat to choke for uptime, or when your volume is too low for pass-through savings to offset the integration work. Five invoices means five support queues when a call fails.

Self-hosted is wrong below roughly 100,000 minutes a month, full stop. The Bearworks analysis shows the fixed infrastructure cost dominating until volume amortizes it. It is also wrong when you lack an engineer who owns GPU capacity planning, model updates, and telephony failover. Self-hosting saves 80-plus percent at a million minutes a month; it costs more than managed at ten thousand.

Premium everything is wrong for internal and back-office lines. Nobody leaving a voicemail for the maintenance line needs a $0.10-per-minute voice.

Common mistakes

• Comparing sticker rates across different bundle definitions. Vapi's $0.05 and Telnyx's $0.05 include different things. Normalize to all-in per-minute cost before comparing.
• Budgeting from the vendor's default estimator. Every estimator defaults to flattering assumptions: average call lengths, cheap models, no hold time. Rebuild the estimate with your call data.
• Benchmarking with a frontier model, shipping with a mini. Or the reverse. Either way your pilot numbers will not survive production.
• Ignoring the resolution rate. A $0.07 per minute agent that resolves half your calls costs more per resolution than a $0.15 agent that resolves nearly all of them.
• Forgetting concurrency. Minute volume tells you the size of the bill; peak concurrent calls tell you whether the platform will even carry your traffic without paid add-ons.
• Signing an annual plan before the pilot. Every major platform offers pay-as-you-go. Run 1,000 real calls before you commit to a minimum spend.

Frequently asked questions

How much does a voice AI agent cost per minute in 2026? Between $0.06 and $0.31 per connected minute for a production agent, depending on the model and voice you choose and how much of the stack you bundle. The five billing layers are telephony, speech-to-text, the LLM, text-to-speech, and orchestration; advertised rates usually cover only one or two of them.

Why is my voice AI bill higher than the advertised rate? The most common causes are billed silence and hold time, per-call rounding to 60-second increments, LLM costs that grow with conversation length, tool-call token overhead, and platform monthly minimums that exceed your usage at low volume.

Is it cheaper to build on Vapi or use an all-in-one platform? Vapi's $0.05 hosting fee covers orchestration only; models, speech, and telephony pass through on top, which lands a typical build at $0.09 to $0.14 per minute. All-in bundles like Deepgram's Voice Agent API at $0.075 or Telnyx's roughly $0.06 stack are often cheaper, but they constrain your model and voice choices. Assembled wins on flexibility, not always on price.

When does self-hosting a voice agent pay off? Above roughly 100,000 connected minutes per month. Below that, the fixed cost of GPU and Kubernetes infrastructure exceeds the per-minute savings. At a million minutes a month, a well-run self-hosted stack can cost 80 to 90 percent less than a managed one, according to published engineering analyses.

Do I pay for silence and hold time on a voice agent? Usually yes. Retell's FAQ states that billing covers the full call duration because the speech-to-text engine stays active throughout. The same is true on most platforms, since STT and telephony meters run for the whole connection.

What does a voice agent cost per resolved call? Divide your total monthly spend by calls resolved without human help. A $9,000 month resolving 24,000 calls costs $0.375 per resolution. This number, not the per-minute rate, is what you should compare against the fully loaded cost of a human-handled call.

Where to go from here

If you have the call data from step one of the estimation procedure, you already know whether voice AI pencils out for your use case and which pricing model fits. We build custom voice agents on whichever stack the math favors, and if your agent needs to do more than talk, our AI employee service puts a voice front-end on an agent that also handles email, scheduling, and back-office work. For the wider cost picture beyond per-minute pricing, see our guides to voice AI implementation costs, what a voice agent costs to run in Kenya, and AI voice assistants for small business.