AI Voice Agent Cost 2026: Real Price Per Minute

An AI voice agent can be advertised at $0.05 per minute and still cost two or three times that amount once it is answering real phone calls. The difference is not necessarily a hidden markup. It is usually the rest of the stack: reasoning, speech recognition, voice generation, carrier time, testing, monitoring, and production support.

In September 2026, the self-serve market spans simple bundled plans, modular orchestration platforms, and model APIs. Their headline rates are not directly comparable.

A realistic starting range is about $0.08 to $0.30 per connected minute before implementation and premium support. Simple, well-optimised configurations can sit below that range; advanced models, premium voices, compliance features, and inefficient calls can push above it.

This guide breaks down the bill and gives worked examples for ElevenAgents, Vapi, Retell, and OpenAI GPT-Live-1.

The Six Parts of AI Voice Agent Cost

A production phone agent can include six separately priced layers.

Cost layer	What it does	How it is usually charged
Platform or session	Coordinates the live conversation	Per connected minute or monthly plan
Speech recognition	Turns caller audio into text or signals	Per minute
Language model	Decides what to say and which action to take	Tokens, messages, or estimated per minute
Voice synthesis	Produces the agent's speech	Characters, tokens, or minutes
Telephony	Provides phone numbers and carries calls	Number rental plus inbound/outbound minutes
Operations	Testing, analytics, storage, compliance, support	Included, usage-based, or monthly add-on

A managed platform may bundle the first four. An orchestration platform may show each one separately. A model API may provide the live voice layer while leaving telephony and application operations to you.

This is why “price per minute” needs a second question: a minute of what, with which components included?

Current Voice Agent Prices at a Glance

The figures below were checked against provider pricing pages on 25 September 2026.

Provider	Published starting point	What remains separate
OpenAI GPT-Live-1	$0.05/live session minute	Backend model, tools, telephony, application operations
ElevenAgents	About $0.08/included or additional minute	LLM usage and external carrier charges
Vapi	$0.05/min hosting; public example totals $0.082–$0.129/min	Carrier, paid support, optional add-ons
Retell AI	$0.07–$0.31/min; current example is $0.11/min	Telephony where used, add-ons, enterprise extras

These prices describe different packages. GPT-Live-1's $0.05 is a front-end voice session. Vapi's $0.05 is an orchestration fee before the selected transcriber, model, and voice. ElevenAgents includes speech-to-text, text-to-speech, knowledge bases, and RAG in its platform minute. Retell's range changes with the chosen language model, voice, and add-ons.

ElevenAgents Pricing

ElevenAgents pricing uses monthly plans with included call minutes:

Plan	Monthly price	Included minutes	Concurrent calls
Free	$0	15	4
Starter	$6	75	6
Creator	$22	275	10
Pro	$99	1,238	20
Scale	$299	3,738	30
Business	$990	12,375	40

Each paid plan works out at approximately $0.08 per included call minute. Additional minutes are $0.08. Calls that exceed the plan's normal concurrency can use burst capacity at $0.16 per minute, up to the allowed burst limit.

The platform includes speech recognition, voice generation, knowledge bases, and RAG. LLM usage is billed separately based on the chosen model. ElevenLabs does not add a telephony fee, but a connected external carrier or SIP provider can charge its own rates.

There is one useful billing distinction: ElevenLabs says voice calls receive a 95% discount for periods of silence longer than 10 seconds. That can matter for interviews, hold-heavy workflows, or calls where users pause while finding information.

If overage is available on the selected account, illustrative base-platform totals are:

Monthly minutes	Example plan calculation	Platform total before LLM/carrier
500	Creator $22 + 225 additional minutes	$40.00
2,000	Pro $99 + 762 additional minutes	$159.96
10,000	Scale $299 + 6,262 additional minutes	$799.96

Those are cost illustrations, not automatic plan recommendations. A higher plan may make sense for concurrency, seats, voices, or operational headroom even when a lower plan plus additional minutes is arithmetically cheaper.

Read our ElevenAgents review for the wider product trade-offs.

Vapi Pricing

Vapi's entry tier charges a $0.05 per-minute hosting fee and passes model costs through separately. Its public pricing calculator currently shows this 1,000-minute example:

Component	1,000-minute estimate
Vapi hosting	$50
Deepgram transcription	about $10
OpenAI intelligence model	$8–$45
ElevenLabs voice	$15–$24
Estimated total	$82–$129

That is approximately $0.082 to $0.129 per minute before a paid support package and any carrier cost. The usage-only tier includes four concurrent calls and 14 days of raw data retention.

The current Core support package is $29 per month. Pro is listed as 10% of the Vapi hosting fee with a $999 monthly minimum and adds a 99% uptime SLA, dedicated Slack support, role-based access control, and longer retention. HIPAA-eligible handling is listed as a $2,000 monthly add-on.

At the calculator's example component range, a simple linear estimate would be:

Monthly minutes	Usage estimate before carrier/support
500	$41–$65
2,000	$164–$258
10,000	$820–$1,290

Actual model use is not perfectly linear. Longer prompts, verbose responses, tool calls, silence, and different voices move the total. Our Vapi review explains why provider flexibility is valuable but makes budgeting more involved.

Retell AI Pricing

Retell publishes a broad $0.07 to $0.31 per-minute pay-as-you-go range. Its current pricing calculator shows a $0.11-per-minute example made up of:

• $0.055 Retell voice infrastructure;
• $0.015 Retell platform voice;
• $0.04 language model;
• $0.00 telephony when custom telephony is selected.

At that example rate, monthly usage would be $55 for 500 minutes, $220 for 2,000 minutes, or $1,100 for 10,000 minutes. At the edges of Retell's published range, 10,000 minutes could instead be $700 to $3,100 before relevant extras.

Retell includes 20 concurrent calls on pay as you go. Optional charges currently include knowledge bases, denoising, safety guardrails, PII removal, AI quality assurance, phone numbers, and additional concurrency.

Its billing rules also show why call design affects cost:

• connected time is measured to the nearest second;
• silence and hold time are billed because the speech engine remains active;
• failed calls are not billed;
• voicemail is billed only while the agent is active;
• the AI-agent fee stops after a transfer, while telephony can continue.

A platform's response to silence, voicemail, and transfers can change the invoice even when two displayed rates look similar.

GPT-Live-1 Pricing

OpenAI's new GPT-Live-1 model costs $0.05 per live session minute, billed per second. It combines listening and speaking in a full-duplex voice layer, but the official model page states that backend Responses calls and tools use their normal separate pricing.

The voice layer alone costs:

Monthly minutes	GPT-Live-1 session cost
500	$25
2,000	$100
10,000	$500

This is attractive for teams that want to build their own product, but it is not equivalent to a finished receptionist platform. Add the backend agent, phone carrier, storage, monitoring, evaluation, and engineering required for the use case.

Our GPT-Live-1 guide explains the architecture and when it makes sense to own that stack.

The Costs Most Forecasts Miss

Telephony and phone numbers

Some pricing calculators show $0 telephony because you selected custom SIP. That means the platform is not charging a carrier fee; it does not mean the telephone network is free. Price inbound and outbound calls in every country you serve, plus number rental, transfers, and recording where applicable.

Concurrency

Monthly minutes describe total volume. Concurrency describes peak volume. Ten thousand minutes spread evenly across a month is different from hundreds of calls arriving after one campaign or outage.

Check the included concurrent calls, the price of extra lines, what callers experience when capacity is exhausted, and whether burst minutes cost more.

LLM and prompt size

The reasoning model can be a small fraction of the bill or one of its largest components. Long system prompts, large knowledge contexts, verbose answers, repeated tool calls, and premium models increase cost.

Use the least expensive model that passes your real task tests. Moving to a cheaper model without testing can save per minute while increasing failed calls and escalations.

Quality assurance

Simulation, automated evaluation, transcript analysis, and call recording can be included or charged separately. They are not optional in practice. An unmonitored agent can give wrong information or mishandle calls at scale before a team notices.

Compliance and data controls

Healthcare, finance, and other sensitive deployments may require BAAs, custom retention, zero-data-retention controls, SSO, role-based access, audit logs, or a dedicated environment. These frequently sit in enterprise contracts or paid add-ons rather than the self-serve minute price.

Implementation and maintenance

A no-code agent still needs call-flow design, a knowledge base, integrations, testing, staff training, and ongoing review. A developer platform adds software engineering and on-call ownership. Include internal time or agency fees in the business case.

Calculate Cost Per Successful Outcome

Cost per minute is useful for infrastructure planning. It is not the final measure of value.

Use this calculation:

Cost per successful outcome = total monthly voice-agent cost ÷ completed outcomes

If a system costs $600 in a month and completes 300 valid bookings, the cost is $2 per booking. If a cheaper system costs $400 but completes only 120, its cost is $3.33 per booking before staff rework.

Track at least:

• connected minutes and average call duration;
• task completion rate;
• human transfer and abandoned-call rates;
• wrong or duplicate tool actions;
• staff time spent correcting calls;
• total cost per booking, resolution, or qualified lead.

A Practical Budgeting Process

1. Export real call data. Estimate connected calls, average duration, peak concurrency, transfer time, voicemail, and seasonal spikes.
2. Define the task. Separate FAQ calls from bookings, payments, support, lead qualification, and regulated conversations.
3. Build the full stack price. Include every component and fixed fee, not just the advertised platform rate.
4. Run a representative pilot. Use real prompts, knowledge, tools, accents, phone lines, and failure scenarios.
5. Measure outcomes. Record completion, transfer, correction, and abandonment rates alongside spend.
6. Model three volumes. Forecast normal, peak, and downside cases—including burst concurrency and longer calls.
7. Add operational cost. Include monitoring, support, compliance, engineering, and content maintenance.

For platform selection, compare our best AI voice agent platforms. For the wider staffing decision, see AI receptionist vs human receptionist.

Verdict

In 2026, an AI voice agent's real price is rarely one number. Expect roughly $0.08 to $0.30 per connected minute for many self-serve production configurations, then add implementation, support, compliance, and any fixed platform costs.

ElevenAgents offers the simplest bundled platform arithmetic. Vapi exposes more component choice and currently shows a competitive $0.082–$0.129 example stack. Retell publishes a wide range with detailed operational add-ons. GPT-Live-1 provides a $0.05 full-duplex voice layer for teams ready to build the rest.

The best-value system is the one that completes the task safely at the lowest cost per successful outcome. Model your actual call mix, test before launch, and treat any headline rate as the first line of the budget—not the last.

Prices checked 25 September 2026 and exclude taxes. Provider rates and plan rules can change; verify the configured total in each provider's current calculator before committing.