How Much Does a Voice AI Agent Cost in 2026? (Real Pricing Models Compared)

Updated on August 14, 2026

TL;DR

Voice AI agent pricing is not standardized across vendors, and that is the root of most buyer confusion.

Voice AI agents cost $0.05 to $0.35 per minute on self-serve platforms. Most production deployments land between $0.12 and $0.25 per minute once every component is billed. Enterprise platforms often do not price per minute at all. They charge per resolution or per business outcome, inside contracts that start in the six figures.

The reason two vendors can quote you $0.05 and $0.30 for what sounds like the same product is that they are usually not selling the same unit. One is selling orchestration. One is selling a completed call. One is selling a resolved ticket. This guide shows you how to normalize those units so you can compare them properly.

Introduction

Voice AI agents have taken the industry by storm in the past two years. New models are released every other week, and they sound more natural with each release. Companies across industries now deploy voice AI agents to qualify leads, schedule appointments, support customers, and issue refunds in real time.

If you are a CX leader or an engineer evaluating voice AI agents, there is a good chance the pricing has confused you. Every vendor in this market uses a different format, and the formats are still evolving. Some platforms charge pay-as-you-go. Some charge per minute. Some charge based on outcomes. Some keep their pricing behind a demo call.

To make evaluation easier, we researched the leading voice AI agent platforms available as of August 2026. For each one, we read the public pricing page, worked through the billing documentation, and checked user reviews on G2. Where a vendor publishes nothing, we say so rather than guessing.

What is a voice AI agent?

A voice AI agent is software that can place or receive a phone call, greet the caller, take in a request, understand it, and resolve it in natural conversation and in real time.

That last part is what separates a voice AI agent from a traditional IVR. An IVR reads a menu and waits for you to press 2. A voice AI agent listens to “I need to move my Thursday appointment to sometime next week,” checks a live calendar, offers real slots, books one, and updates your CRM.

Modern voice AI agents can reason, respond, and take action through integrations with your CRM and internal databases. They also know when to stop and hand the call to a human.

Every voice AI agent runs on five core layers: telephony, automatic speech recognition, reasoning with LLMs, orchestration, and text-to-speech. Understanding these five layers is the key to reading any pricing page, because every vendor bundles a different subset of them into its headline rate.

How a voice AI agent works: the 5 layers

Every voice AI call, on every platform, runs the same pipeline. A caller speaks. The audio travels over a phone network, gets transcribed, gets reasoned over, gets synthesized back into speech, and travels back.

Five layers. Every pricing model is just a different way of packaging them.

Layer 1: Telephony and carrier

This is the entry point for every voice AI agent, and the one layer nobody can abstract away. The call has to physically exist on the phone network.

Telephony is what allows your agent to answer a public phone number. The voice AI platform connects to a telephony provider through an API, and that provider handles the actual call.

The major providers are Twilio, Telnyx, Aircall, and Ozonetel. Some platforms let you choose. Kommunicate, for example, supports Twilio and Ozonetel.

Provider	US outbound	US inbound (local)	Notes
Twilio	~$0.013–0.014/min	~$0.0085/min	Toll-free inbound ~$0.022/min; number ~$1.15/mo
Twilio BYOC trunking	~$0.004/min	~$0.004/min	Your existing carrier still bills you separately
Telnyx	~$0.005/min	~$0.001–0.0032/min	Call Control API adds ~$0.002/min
Ozonetel	quote-based	quote-based	Strong India coverage, built for contact centers

The per-minute rate is what everyone compares, but the line items underneath it are what actually move your bill. Watch for phone number rental, concurrency channels, toll-free premiums, A2P registration, and international rates that can run three to five times the US rate. Telnyx, for instance, sells inbound channels at roughly $12 per month each for the first ten.

Call recording is billed separately again, at roughly $0.0025 per minute to record plus $0.0025 per minute to store on Twilio.

Layer 2: Automatic speech recognition (ASR)

ASR, also called speech-to-text or STT, transcribes what the caller says into text so the language model can understand it.

This is the cheapest layer, and it has been getting cheaper fast:

• Deepgram Nova-3: $0.0043/min pre-recorded, $0.0077/min streaming
• AssemblyAI streaming: ~$0.0025/min
• Cartesia Ink-Whisper: ~$0.00217/min
• Self-hosted open models: effectively free, plus GPU cost

At under a cent a minute, ASR is rarely worth optimizing for price. It is very much worth optimizing for accuracy, and this is the part vendors rarely explain. Cheap transcription that mishears the caller forces the agent to ask again. Every repeat turn costs you LLM tokens, TTS characters, and telephony minutes. A 5% accuracy improvement can pay for itself several times over in the layers above it.

This matters most outside English. Global ASR engines underperform on Indian regional languages, Southeast Asian languages, and heavily accented English. That is why India-first vendors like Bolna and Sarvam build their own transcription routing instead of defaulting to a single provider.

One add-on to watch: speaker diarization on Deepgram, which labels who said what, is a paid extra at roughly $0.002 per minute. It is not included by default.

Layer 3: Reasoning with NLP and LLMs

This is the brain of the agent. It interprets intent, holds the conversational context, decides what to say, and decides when to call a tool such as your CRM. It then passes the response to the text-to-speech layer.

OpenAI, Anthropic, and Google lead this segment with their GPT, Claude, and Gemini models. The stronger the model, the better it handles complex or ambiguous queries.

Retell AI publishes its per-minute LLM rates, which gives an unusually clear view of how much model choice matters:

Model	Cost per minute
GPT 5 nano	$0.003
GPT 5 mini	$0.012
Gemini 3.1 Flash Lite	$0.014
Claude 4.5 Haiku	$0.025
GPT 4.1	$0.045
Claude 4.6 Sonnet	$0.08
GPT 5.5	$0.16

That is a 50x spread on the same platform, for the same call. Two things follow from it.

First, model choice is the single biggest cost lever you control. Swapping a frontier model for a smaller one on simple, scripted calls can cut your LLM line by 80% or more without the caller noticing.

Second, LLM cost scales with the number of conversational turns and the length of your context, not with clock time. A rambling six-turn call costs far more than a crisp six-turn call of the same duration. Tightening your prompts is a cost exercise, not just a quality one.

Layer 4: Orchestration

Orchestration is the layer that turns three separate APIs into something that behaves like an agent. It handles turn-taking and interruptions, keeps track of conversation state, manages retries, calls functions in your CRM and backend systems, runs your human handoff logic, and produces the logs and dashboards you review afterward.

This is the layer most platforms charge for, and often the only one they advertise.

Bolna is a clear example of the pure orchestration model. You pay for the orchestration layer alone and bring your own API keys for ASR, reasoning, TTS, and telephony.

Layer 5: Text-to-speech (TTS)

TTS turns the text response back into audio. It is usually the most expensive AI layer, and it is billed per character rather than per minute, so you only pay while the agent is actually speaking.

Tier	Rate	Examples
Commodity cloud	~$4 / 1M chars	Amazon Polly Standard, Google Standard
Mainstream neural	$12–30 / 1M chars	Deepgram Aura-2 ($30/1M), Azure Neural
Premium realtime	~$50 / 1M chars	ElevenLabs Flash/Turbo ($0.05/1k chars)
Ultra-realistic	$100–160 / 1M chars	ElevenLabs Multilingual v2/v3 ($0.10/1k chars)

An agent that talks for roughly half a call generates about 450 characters per call-minute. That works out to about $0.007 per minute on mainstream neural voices and about $0.045 per minute on a flagship model, a six-fold difference.

Treat this as a brand decision rather than a technical one. A premium voice is worth paying for on inbound calls, where the agent represents your company to a customer who already has a problem. It is largely wasted on outbound appointment reminders.

Why voice AI agents are growing fast in 2026

AI voice agents answered roughly 14% of inbound small-business calls in late 2025, up from about 2% a year earlier. That is a sevenfold jump in twelve months. Here is what is driving it.

24/7 availability without night-shift cost: A call at 2am gets the same answer as a call at 2pm. For clinics, insurers, and e-commerce teams, after-hours volume was previously either lost or routed to an expensive answering service.

Concurrency: One human agent handles one call. One voice AI agent handles ten simultaneous calls out of the box on most platforms, and scales to hundreds on enterprise tiers. Peak-hour queues stop being a staffing problem and become a configuration setting.

Appointment booking and rescheduling: Clinics, dental practices, salons, and service businesses run on scheduling. Agents check real availability, then book, reschedule, cancel, and confirm. This is the highest-ROI use case in the category because the workflow is bounded and the value of each completed action is easy to measure.

CRM and backend integration: Voice AI agents do more than talk. They read order status from your commerce backend, write call outcomes to Salesforce or HubSpot, trigger refunds, and update ticket records. This is what turns a demo into a deployment.

Lead qualification at speed: Agents call new leads within seconds of a form submission, qualify them against your criteria, and push structured data into the CRM. Speed-to-lead like this used to require a large SDR bench.

Human handoff when it matters: The best deployments do not aim for 100% automation. They handle the routine 30% to 50% of calls and escalate the rest with the full transcript and context attached, so the human does not start from zero.

The economics finally work: Human-handled calls are commonly cost at up to $12 each. An AI-handled call runs $0.30 to $0.50. Even at partial automation rates, that arithmetic is hard to argue with.

The 7 pricing structures vendors use today

This is the section that resolves the confusion we described at the start. There are seven distinct pricing structures in the market. Each one is a rational commercial choice aimed at a specific type of buyer.

The list below runs from structures where the buyer carries most of the risk to structures where the vendor does.

1. Bring-your-own-key (BYOK)

How it works: The platform charges for orchestration only. You supply your own API keys for ASR, LLM, TTS, and telephony, and pay each provider directly.

Who uses it: Bolna and Vapi.

Best for: Teams with engineering capacity and a specific reason to control the stack, such as a latency target or a language requirement.

The catch: Four or five separate invoices, no single vendor accountable when latency degrades, and a headline rate that understates your true cost by two and a half to five times.

2. Fixed all-inclusive per minute

How it works: One published rate covers every layer. What you see is what you are invoiced.

Who uses it: Kommunicate ($0.06/min), Bland ($0.11–0.14/min), Ringg (from ~$0.08/min).

Best for: Buyers who need to forecast. Finance teams that want one line item instead of five. Anyone running a pilot who needs a defensible number for budget approval.

The catch: You pay the bundler’s margin, and you lose component-level control. You generally cannot swap in a cheaper model the week it ships.

3. Subscription plus bundled minutes

How it works: A monthly fee either includes an allowance of minutes or unlocks a lower per-minute rate.

Who uses it: Bland’s tier structure, ElevenLabs’ credit plans, Synthflow’s Enterprise tier.

The catch: Under-utilization quietly doubles your effective rate, and the arithmetic is worth doing explicitly.

Bland restructured its pricing in December 2025. Critically, the plan fee includes zero minutes. You pay the subscription, then pay usage on top:

Plan	Monthly fee	Per-minute rate
Start	$0	$0.14
Build	$299	$0.12
Scale	$499	$0.11

Run the break-even and something counterintuitive appears. The Build plan only beats the free Start plan above roughly 14,950 minutes per month. The Scale plan only beats Build above 20,000 minutes per month.

At 5,000 minutes a month, the free tier costs $700 while the $499 Scale plan costs $1,049. The discount plan is 50% more expensive.

Do this calculation for any tiered vendor before you upgrade. The upgrade prompt in the dashboard will not do it for you.

4. Per connected minute

How it works: You are billed only for minutes where the call actually connects, not for dial attempts, ring time, or dead numbers.

Best for: Outbound operations with low connect rates, such as collections or cold outreach. If only 30% of your dials connect, this model is dramatically better than paying per dialed minute.

The catch: The definition of “connected” is the whole ballgame, and it is rarely published. Does voicemail count? A two-second hangup? An answered call where nobody speaks? We looked for a precise published definition across the vendors marketing this model and could not find one on a public pricing page. That should be your first question on the sales call.

5. Per resolution or per conversation

How it works: You pay for each conversation the AI fully resolves. Escalations to a human cost less, or nothing.

Who uses it: Parloa, PolyAI (reportedly more than 40% of its customers), Fin, Zendesk AI Agents.

The catch: “Resolution” is defined by the vendor. Does it still count if the customer calls back about the same issue two days later? Per-resolution pricing also looks affordable at pilot scale and becomes unpredictable at volume, because your cost is now tied to the AI’s success rate.

6. Outcome-based

How it works: You pay for a business result, such as a retained subscriber, an originated mortgage, or a processed claim, rather than for the conversation that produced it.

Who uses it: Sierra is the clearest example.

The argument for it: Pricing should follow how much agency the software has and how cleanly the result can be attributed to it. Where the agent is autonomous and the outcome is attributable, outcome pricing fits. The underlying bet is that intelligence keeps getting cheaper, so anything priced on top of compute compresses over time, while outcomes do not.

The catch: Sierra itself is candid that this model is harder to run operationally, contractually, and in accounting terms. Attribution disputes are structural rather than occasional. In practice it also means no published price, long procurement cycles, and no access below a certain scale.

7. Enterprise custom contract

How it works: An annual contract bundling license, volume commitment, integration work, and managed delivery.

Who uses it: PolyAI (six-figure annual contracts are typical), Parloa (reported entry around $300k per year, unconfirmed), Cresta, Cognigy, Decagon.

The catch: Multi-month deployments, implementation that depends on certified partners, and pricing opacity that makes genuine comparison nearly impossible.

10 best voice AI agent platforms and their pricing in 2026

This list is ordered by pricing transparency and how easily you can forecast your bill before signing. It is not a ranking of voice quality, latency, or feature depth. Those vary by use case.

Platform	Pricing model	Entry price	Platform fee	ASR/STT	LLM	TTS	Telephony	G2 Rating
Kommunicate	All-inclusive	$0.06/min	$40/month	Included	Included	Included	Twilio / Ozonetel, billed by carrier	4.8/5
Retell AI	Pay as you go / BYOK	$0.07/min	$0.055/min	Included	$0.003–0.16/min	$0.015–0.040/min	$0.015/min, or free with own SIP	4.8/5
Vapi	Component / BYOK	$0.05/min + components	$0.05/min	At cost or BYOK	At cost or BYOK	At cost or BYOK	Your provider	4.2/5
Bland AI	All-inclusive	$0.14/min (free tier)	Bundled	Included	Included	Included	Included	5/5
Bolna	Orchestration + BYOK	6.00¢/min (₹5.52)	~$0.02/min	BYOK	BYOK	BYOK	Twilio, Plivo, Exotel	No reviews
Ringg AI	All-inclusive	From ~$0.08/min	Bundled	Included	Included	Included	Included	4.8/5
Synthflow	Usage-based + Enterprise	Demo required	Not published	Included	Included	ElevenLabs	Synthflow or BYO	4.5/5
ElevenAgents	Subscription + per minute	$0.10/min + LLM	$0.08–0.10/min	Included	Billed separately	Included	Your carrier	No reviews
PolyAI	Per minute or per resolution	Six figures per year	Not published	Included	Included	Included	Included	5/5
Sierra	Outcome-based	Custom	Not published	Included	Included	Included	Included	4.4/5

Verified against vendor pricing pages, August 2026. “Not published” means the vendor does not disclose that component publicly. Retell, Kommunicate, and Vapi are all model-agnostic across OpenAI, Anthropic, and Gemini.

1. Kommunicate

Kommunicate is a customer service automation platform that helps CX teams automate support across voice, chat, WhatsApp, and email. Founded in 2020, it was one of the first Indian AI companies selected for Google’s AI-First Accelerator and now serves more than 500 customers globally.

Kommunicate charges a flat $0.06 per minute for Voice AI on both the Starter ($40 per month) and Professional ($200 per month) plans, with custom rates at Enterprise. There is no per-model LLM tier, no separate TTS meter, and no voice-infrastructure line. The $0.06 covers the AI stack. Telephony connects through Twilio or Ozonetel and is billed by your carrier, so you can bring your own.

Because voice sits on the same plan as chat, email ticketing, and live chat, it works as part of a broader support stack rather than a standalone tool. For a CX leader who needs one number to take into a budget meeting, that predictability is the main draw.

Best for: CX leaders who need a defensible, predictable cost, and teams running voice alongside chat and email rather than in isolation.

Trade-off: Less component-level control than a BYOK platform. You cannot swap in a cheaper model yourself.

2. Retell AI

Founded in 2023, Retell AI helps teams build, deploy, and manage voice agents that sound human and execute real tasks. Retell supports 31+ languages and claims to respond within 600 milliseconds. 

Retell has the most complete published rate card in the market. It breaks out every component: $0.055 per minute for voice infrastructure, TTS from $0.015 per minute (ElevenLabs voices at $0.040), per-model LLM rates from $0.003 to $0.16 per minute, and telephony at $0.015 per minute or free if you bring your own SIP trunk. Retell allows you to build the AI agent on top of your choice of large language models (LLMs).

Its published band is $0.07 to $0.31 per minute, and both ends are real. The floor is a nano model with platform voices and your own SIP trunk. The ceiling is GPT 5.5 with ElevenLabs voices and Retell’s telephony. Billing is per second, with no rounding up per call.

Best for: Technical teams who want to model their costs precisely before committing.

Trade-off: You have to actively manage the component stack to keep costs where you want them.

3. Vapi

Founded in 2020, Vapi is a cloud platform that lets developers add voice calling and messaging to their own applications.

Vapi has the lowest published platform fee in the category at $0.05 per minute. Model costs pass through at cost, or drop to zero if you supply your own API keys. The plan includes 10 concurrent calls, with additional lines at $10 per line per month.

Watch the compliance add-ons. HIPAA is $2,000 per month and Zero Data Retention is $1,000 per month, which is $36,000 a year before a single call connects.

Best for: Engineering teams optimizing their own stack.

Trade-off: Four to five separate invoices, and no single vendor accountable for latency.

4. Bland AI

Founded in 2023, Bland builds voice agents that handle phone calls at volume. It runs its own models for transcription, language, and voice in-house rather than reselling third-party APIs.

That control shows in the output. Bland agents are among the most realistic in the category, adjusting pace based on how the caller sounds.

Pricing is genuinely all-inclusive across three tiers: $0.14 per minute on the free Start plan, $0.12 at $299 per month, and $0.11 at $499 per month. Bland charges a $0.015 minimum on outbound calls that fail or end within about ten seconds.

Best for: High-volume outbound campaigns above roughly 15,000 minutes per month.

Trade-off: A closed stack. You use Bland’s voices and models.

5. Bolna

Bolna is built specifically for Indian languages, enabling businesses to automate calls with multilingual voice agents. It supports 20+ languages and 200+ accents and dialects, including Hinglish, Hindi, Tamil, and Telugu.

Pricing is published at 6.¢ per minute (₹5.52), falling to 4.51¢ per minute on higher volume tiers. The Pilot plan gives 10,000 minutes at 6¢ plus a 20% bonus, an effective 5¢ per minute. Billing runs in 30-second pulses, and BYOK is supported for LLM, TTS, and ASR. Native integrations with Exotel and Plivo make the telephony side straightforward in India.

Best for: Indian and South Asian outbound calling at volume.

Trade-off: Less relevant outside South Asian markets.

6. Ringg AI

Ringg AI builds and runs AI agents across voice, WhatsApp, chat, web, and email.

It positions an all-in bundled rate from around $0.08 per minute, with sub-400ms latency on South Asian language models. The public pricing page states usage-based pricing with a monthly minimum commitment, so the real number is quote-based.

Best for: South Asian deployments that want a single bundled invoice.

Trade-off: You cannot model your cost from the website. You have to talk to sales.

7. Synthflow

Synthflow offers an end-to-end voice AI platform with in-house telephony. It has one of the strongest visual flow builders in the category, with 200+ integrations and 30+ languages.

One important note: Synthflow restructured its pricing in 2026, retiring the old $29 to $899 tiers in favor of pay-as-you-go plus an Enterprise plan starting around 10,000 minutes per month. Any figure you find quoting those old plans is out of date. Concurrency is the cost to watch here, with 5 calls included and additional slots at $20 per month up to 50.

Best for: Teams with no engineering resource who need a visual builder.

Trade-off: Concurrency costs escalate faster than on most competitors.

8. ElevenLabs Agents

ElevenLabs is an AI research company that builds human-like voice technology. Its ElevenAgents lets you create voice agents from a dashboard with or without code, while the API gives you the raw models to build your own.

Voice quality remains the category benchmark. Agents start at $0.08 per minute on annual Business plans and $0.10 per minute on Creator and Pro.

The critical caveat is that LLM costs are billed separately, so the headline rate is not your full cost. Burst concurrency runs at double the standard rate.

Best for: Consumer-facing inbound calls where voice quality is the product.

Trade-off: You still have to source and pay for the reasoning layer yourself.

9. PolyAI

PolyAI builds enterprise conversational AI designed to understand customers regardless of how they phrase things. It reports resolving over 50% of calls automatically.

Six-figure annual contracts are typical. Pricing is per minute by default, though reportedly more than 40% of its customers have now moved to per-resolution. PolyAI has deep expertise with complex enterprises still running mainframes and on-premise systems.

Best for: Large contact centers that need managed delivery and legacy integration.

Trade-off: No published pricing and a long procurement cycle.

10. Sierra

Sierra is an enterprise voice AI company. Its Ghostwriter product lets CX teams build and deploy agents without coding, and it typically works through forward-deployed engineers who spend months with a client understanding their operation before deploying custom agents.

Sierra publishes no pricing and charges for business outcomes rather than conversations. More than 40% of the Fortune 50 are customers, and it holds FedRAMP High, PCI DSS Level 1, and HIPAA certification.

Best for: Large enterprises where a single interaction carries real revenue weight.

Trade-off: Inaccessible below a certain scale, and attribution needs to be negotiated carefully.

The extra costs nobody quotes you

The rate on the pricing page rarely includes the operational costs that show up later: storage for call recordings and how long you can access them, how many calls your agent can handle simultaneously, integrations with your CRM, and the human hours needed to set the whole thing up.

These typically add 15% to 30% to a naive estimate, and sometimes considerably more.

Call recording and storage

Twilio charges roughly $0.0025 per minute to record and another $0.0025 per minute to store. Post-call transcription for analytics is billed separately again, up to $0.05 per minute on some providers, and it is distinct from the live ASR you have already paid for.

Check the retention period too. A platform that keeps recordings for 30 days and one that keeps them for a year are not offering the same thing, and extending retention is usually a paid upgrade.

Concurrent calls

This is a monthly fee with nothing to do with minutes, and it surprises buyers more than anything else on this list.

Platform	Included concurrency	Additional
Retell	20	$8/slot/month
Vapi	10	$10/line/month
Synthflow	5	$20/slot/month, up to 50
Bolna	Up to 100 (Pilot plan)	Plan-dependent
Telnyx (carrier)	—	~$12/channel/month, first 10

At 30 concurrent calls, that is $80 per month on Retell and $500 per month on Synthflow. Model your peak, not your average. A Black Friday queue costs far more in lost orders than you would ever save on the per-minute rate.

AI summaries, QA, and per-feature metering

Some platforms meter individual features on top of the base rate. Retell publishes several that are easy to miss: knowledge base at +$0.005 per minute, PII removal at +$0.01, advanced denoising at +$0.005, safety guardrails at +$0.005, AI quality assurance at $0.10 per minute, and branded caller ID at +$0.10 per outbound call.

Turn on four of those and you have added $0.12 per minute, which is more than the base rate itself.

Plan-based platforms handle this differently. On Kommunicate, AI Summary and auto-resolve are bundled into the Professional tier rather than metered per use. That is a real trade-off rather than a straight win: less granular control, more predictable billing.

Compliance add-ons

Vapi charges $2,000 per month for HIPAA and $1,000 per month for Zero Data Retention. Most other platforms gate HIPAA behind enterprise tiers with custom pricing. Budget this as a fixed annual cost rather than a per-minute one.

Standing monthly charges

Phone numbers ($2 per month on Retell, around $1.15 on Twilio), verified numbers ($10 per month), knowledge bases ($8 per month each beyond the first ten), SMS capability ($20 per month), and extra AI agents ($20 to $30 per month on seat-based platforms).

Human hours: the cost that breaks budgets

This appears on no pricing page and is frequently the largest line item in year one.

• Self-serve DIY: 40 to 80 hours of setup across prompt design, testing, and integration. At $75 to $150 an hour of internal time, that is $3,000 to $12,000 before your first production call.
• Agency or managed: typically $500 to $3,000 for one-time setup, plus a $300 to $1,500 monthly retainer at SMB scale.
• Enterprise forward-deployed engineering: vendors like Sierra, Parloa, and PolyAI embed engineers in the deployment. Sometimes this is bundled, sometimes billed as professional services, sometimes delivered by certified partners. Parloa deployments reportedly run one to three months.
• Ongoing supervision: 0.5 to 1 full-time equivalent for a basic deployment, rising to 2 or 3 for complex enterprise rollouts.

The cheapest platform is frequently the most expensive deployment. Any comparison that stops at per-minute rates is comparing about 60% of your total cost.

Billing mechanics that quietly inflate invoices

• Increments. Per-second, 30-second pulse, or per-minute rounding. On a fleet of 40-second calls, rounding up to the minute inflates your bill by 50% at an identical headline rate.
• Silence. Retell states plainly that billing covers the full call duration including silence, because the ASR engine stays active throughout. Most platforms work the same way. Ask about auto-hangup-on-silence defaults.
• Transfers. Retell stops the AI fee and continues telephony when a call transfers. Bland charges $0.03 to $0.04 per minute during transfers on its own numbers, and nothing if you use your own Twilio account. Some vendors bill both legs.
• Repricing risk. Bland’s December 2025 change reportedly added 25% to 55% to existing customers’ bills. Ask what notice period and price protection your contract carries.

How to compare voice AI agent pricing: a checklist

Before you sign anything, get answers to these ten questions in writing.

1. What is the billing increment: per second, or rounded up per call?
2. Am I billed during silence, hold time, and IVR navigation?
3. Is a call that reaches voicemail billed? For the full duration, or only while the agent speaks?
4. When a call transfers to a human, which meters keep running?
5. What is your precise definition of a “connected minute” or a “resolution”?
6. How many concurrent calls are included, and what do additional ones cost per month?
7. Which features are metered separately from the base rate? Can I see the full add-on list?
8. What are the standing monthly charges for numbers, knowledge bases, SMS, and compliance?
9. Is implementation included, billed separately, or delivered by a third party?
10. What notice do I get before a price change, and is there price protection in the contract?

Frequently asked questions