AI voice agent pricing in 2026: flat hourly rate vs stacked per-minute costs

AI voice agent pricing usually lands between a few cents and a few tens of cents per minute. Where you land depends on the pricing model: per-minute platforms with pass-through model fees, per-layer stacks you assemble, token-metered speech models, or a flat all-in hourly rate such as AssemblyAI's Voice Agent API at $4.50/hr ($0.075/min), billed per second. On public price lists, platform hosting starts at $0.05 a minute before any models are added, and all-in platform rates run from $0.07 to past $0.30 a minute depending on what you pick.

The rate card is the easy part. The hard part is that most of these models bill on several meters at once, and each meter moves when you change the agent. A longer system prompt, a chattier voice, a new tool call: every one of those changes the price of a minute.

So this guide does something the rate cards don't. It lays the four pricing models side by side, opens up a minute of agent time to see what's inside it, and prices the same 5-minute call three ways with the arithmetic shown. All public prices were checked on October 20, 2026.

The four ways voice agents are priced

Competitor prices from public pricing pages as of October 20, 2026: Vapi, Retell, OpenAI. Hosted agents platforms such as ElevenLabs Agents mix models: plans with included call minutes, $0.08 per additional minute, and LLM usage billed separately.

None of these models is wrong. They answer different questions. A platform fee buys speed to a working agent. A self-built stack buys control. Token pricing buys access to a multimodal model. A flat rate buys a number you can multiply.

What's inside a minute of agent time

Every voice agent does the same work each minute, whoever bills for it. The question is how many invoices that work lands on.

One line to be clear about: telephony is still yours to pay. On the Voice Agent API, your Twilio account bills the carrier minutes and AssemblyAI bills the session. There's just no markup in between. The full rate card, including transcription-only rates, is on the pricing page, and we cover the speech layer on its own in speech-to-text pricing and effective cost.

Pricing one 5-minute call three ways

Same call, three pricing models. Here's the arithmetic.

1. Flat hourly rate

5 minutes × $0.075 = $0.375, plus your carrier minutes. That's the whole calculation. It doesn't change if the caller rambles or the agent calls three tools.

2. Per-layer stack

Price the layers you can price exactly, and write the rest as variables.

• Speech-to-text: Universal-3.6 Pro Realtime at $0.45/hr is $0.0075/min. 5 × $0.0075 = $0.0375. (On standalone STT, prompting at +$0.05/hr and Voice Focus at +$0.10/hr are add-ons; both are included in the Voice Agent API.)
• LLM: (input tokens × input rate + output tokens × output rate), summed over every turn. With GPT-5 mini on LLM Gateway at list prices of $0.25 input and $2.00 output per 1M tokens (in-region routing adds 10%), assume 12 agent turns, an average of 2,500 tokens of prompt and history re-sent per turn, and 60 tokens per reply. That's 30,000 × $0.25/1M = $0.0075, plus 720 × $2.00/1M = $0.0014, about $0.009.
• Voice generation: characters or seconds spoken × your voice vendor's rate.
• Turn detection, hosting, and orchestration: your infrastructure cost per minute, or a platform fee.
• Telephony: your carrier's rate.

Speech and reasoning come to about $0.046 before you add the voice, the infrastructure, and the phone line. Change the prompt length or the turn count and the LLM line moves with it.

3. Token-metered speech model

Here the method matters more than any single number. With OpenAI's gpt-realtime-2.1, you pay $32 per 1M audio input tokens for what the caller says, $64 per 1M audio output tokens for what the agent says, plus text tokens for instructions and the context re-sent each turn (cached input is far cheaper at $0.40 per 1M). Converting minutes to tokens depends on how the conversation runs, so the reliable way to price it is to measure a sample of your own calls. GPT-Live-1 changes the shape: a $0.05/min session fee (5 minutes = $0.25), plus backend model and tool usage on top. Our AssemblyAI Voice Agent API vs OpenAI Realtime comparison covers the non-price differences.

The same call at public list rates

So the stack can come out cheaper or more expensive than $0.375. What changes is how many variables you have to forecast. For a closer look at how the platform and framework options differ beyond price, see Vapi vs Pipecat vs LiveKit.

Monthly math at three volumes

Per-call math is useful. Finance wants the month.

• 1,000 calls a day at 4 minutes each: 4,000 minutes × $0.075 = $300 a day, or about $9,000 for a 30-day month.
• 10,000 hours a month: 10,000 × $4.50 = $45,000 a month.
• A 10x spike: 40,000 minutes a day is $3,000 a day. That's ten times the bill and nothing else: no overage tier and no concurrency fee. Accounts do have a concurrent-session limit, which scales on request, so plan the ramp with our voice agent launch runbook.

Add your carrier minutes to each line. Everything else is already in the number.

"The capabilities AssemblyAI enables us to build help businesses market confidently while saving time and money. It's powerful, almost magical to see it work."

— Ryan Johnson, Chief Product Officer, CallRail

What makes per-minute and token bills drift

These are the levers that move a voice agent bill after launch. Most of them move when you improve the agent.

• Prompt and history length. Many stacks re-send the system prompt and conversation history every turn, so long calls get more expensive per minute as they go.
• Agent talk ratio. Output tokens and voice generation usually cost more than input. A chattier agent costs more.
• Prompt caching. Cached input can be a fraction of the uncached rate, so cache hit rate swings the LLM line.
• Tool-call turns. Each tool call adds a round trip of tokens. See how to cut LLM costs on voice workloads.
• Platform minimums. Some support packages carry a monthly minimum, such as Vapi's Pro package at $999 a month.
• Concurrency lines and burst pricing. Retell charges $8 per extra concurrent call per month beyond 20; ElevenLabs bills burst traffic above your plan's limit at twice the rate.
• Silence and hold time. Per-minute models generally bill connected time, silence included. The Voice Agent API does too, which is why call length is the number to manage.
• The resume window. On the Voice Agent API, closing the socket without sending session.end keeps the session resumable, and billable, for up to 30 more seconds. Send session.end when a call is over.

Line items buyers miss

Use this as a checklist when you read any voice agent quote, ours included:

• Concurrency caps, per-line fees, and what happens at the cap
• Per-agent or per-seat subscriptions
• Telephony markup on top of the carrier rate
• Billing granularity: per second, or rounded up to the minute
• Idle and connected time: does an open socket with no audio still bill?
• Overage rates on committed contracts
• Data residency surcharges (some LLM and speech providers charge 10% more for regional endpoints, so ask whether yours does)
• Support tiers and their minimums

The bill is part of the test. Run your pricing assumptions through the same production-readiness checklist you use for latency and errors.

When a per-component stack is the better deal

A flat rate isn't always cheapest. A per-component stack can win when:

• Your calls are short and scripted, so the LLM line stays small.
• You already run inference infrastructure and on-call for it.
• One layer is your differentiator, such as a fine-tuned LLM or a custom voice.
• Your volume is high enough to negotiate four contracts, and you have someone to manage them.

If that's you, put the speech layer first. Universal-3.6 Pro Realtime drops into LiveKit and Pipecat at $0.45/hr. Just price the cost that isn't on any invoice too: the integration, debugging, and incident triage across vendors. We cover that in the engineering cost of a multi-vendor stack.

How to estimate your own AI voice agent cost in 10 minutes

Start with one formula:

monthly cost = calls per month × average connected minutes × rate per minute

Then divide by the calls your agent actually resolves to get cost per completed task, which is the number that matters for your margins. Measure average connected minutes from a pilot, not from your script. Real callers pause, repeat themselves, and ask follow-ups the script didn't plan for.

RATE_PER_HOUR = 4.50  # Voice Agent API, billed per second


def monthly_cost(calls_per_day, avg_minutes, days=30, drops_without_end=0):
    # drops_without_end: calls closed without session.end (up to 30 s more each)
    seconds = calls_per_day * days * avg_minutes * 60 + drops_without_end * 30
    return seconds / 3600 * RATE_PER_HOUR


def cost_per_completed_task(monthly, resolved_calls):
    return monthly / resolved_calls


monthly = monthly_cost(1000, 4)
print(monthly)  # 9000.0
print(round(cost_per_completed_task(monthly, 24000), 3))  # 0.375

In this example, 30,000 calls a month at 4 minutes each costs $9,000. If the agent resolves 24,000 of them, each completed task costs $0.375. Improve the resolution rate and that number falls even though the per-minute rate didn't move.

Choosing a pricing model

• If you need a working phone agent this week and your flow is standard, a per-minute platform gets you there fastest. Budget for the models on top.
• If one layer is your product, assemble a stack and own that layer. Budget engineering time as a cost line.
• If your agent is multimodal, a token-metered model fits. Price it from measured calls, not estimates.
• If finance needs to know what a minute costs before you scale, choose a flat all-in rate.

Price is one criterion. For accuracy, turn-taking, and control, see our comparison of the best voice agent APIs.

Price the conversation, not the components

Voice agent pricing is drifting away from components and toward the conversation. Five years ago you bought minutes of transcription. Today you're buying a whole exchange: listening, reasoning, speaking, turn-taking, and the infrastructure underneath.

That changes the comparison. A cheaper meter on one layer doesn't help if three other meters move every time you ship. The teams that price their product well are the ones who can say what a completed call costs, today and after next month's prompt change. Pick the model that lets you answer that.

Frequently asked questions

How much does an AI voice agent cost per minute?

An AI voice agent typically costs a few cents to a few tens of cents per minute, depending on the pricing model and the models you choose. Public platform hosting fees start at $0.05 per minute before model costs, all-in platform rates run from about $0.07 to $0.31 per minute, and AssemblyAI's Voice Agent API is $0.075 per minute all in, billed per second. Carrier minutes are extra on every option.

How expensive are AI agents to run?

Running a voice AI agent costs roughly the number of connected minutes times your per-minute rate, plus carrier fees. At AssemblyAI's $0.075 per minute, 1,000 four-minute calls a day costs $300 a day, or about $9,000 a month, and 10,000 hours a month costs $45,000. On per-component stacks, the same volume varies with prompt length, turn count, and talk time.

Voice agent API pricing comparison: which is most cost-effective per hour at scale?

At scale, the most cost-effective voice agent API is usually the one with the fewest metered lines you have to forecast. AssemblyAI's Voice Agent API is a flat $4.50 per hour, billed per second, with speech, LLM, and voice included and no concurrency fees. Token-metered and per-component options can be cheaper on short scripted calls, but their hourly cost changes with conversation shape, so compare on a sample of your own calls.

How does flat-rate voice agent pricing compare to per-component STT, LLM, and TTS billing?

Flat-rate voice agent pricing charges one rate per second of connected time, while per-component billing meters speech, LLM tokens, voice, and telephony separately. Flat rates are forecastable because only call length moves the bill. Per-component billing can come out lower or higher, and it shifts every time you change the prompt, the model, or how much the agent talks.

What's the total cost of ownership for a voice agent handling 10,000 hours a month?

On AssemblyAI's Voice Agent API, 10,000 hours a month costs $45,000, plus your carrier's per-minute rate. That covers speech-to-text, the LLM, voice generation, turn detection, recordings, and hosting, with no per-layer add-ons, concurrency fees, or per-agent subscriptions. An assembled stack at the same volume adds engineering and infrastructure costs that don't appear on any single invoice.

Do you need to pay for voice AI?

You can start building voice AI without paying: AssemblyAI gives new accounts a free account with API credit and no credit card required. After that, the Voice Agent API bills per second at $4.50 per hour, with volume discounts available for large workloads. Other providers offer free credits or plan minutes, then bill by minute or token.