AI voice agent pricing: a practical guide to total cost

What AI voice agents cost in 2026

The useful number is your total cost per successful outcome. Start with every billable component and fixed operating cost, then divide the total by resolved calls, booked appointments, qualified leads, or whichever result the agent must deliver.

Published starting rates measure different layers. Speech-to-text (STT), text-to-speech (TTS), large language model (LLM) usage, and the phone call itself may sit outside a platform's headline minute. Our Growth plan starts at $0.08 per minute before internet telephony (VoIP) and LLM usage.

These are USD list-price inputs, not equivalent all-in offers. Vapi's $0.05 hosting minute and Telnyx's $0.05 voice-engine minute include different components. Bland bundles the AI components while keeping carrier cost separate. Synthflow prices an enterprise deployment. The lowest visible number does not establish the lowest total cost.

Why the advertised minute is not the total cost

An AI voice agent creates costs from the first dial or inbound connection through the business outcome. The exact boundary depends on the architecture and contract.

Call attempts and telephony

Telephony depends on direction, destination, number type, and the number of live call legs. An outbound attempt may add answering-machine detection, branded calling, or dialer fees before a useful conversation occurs. A transfer can create another phone leg or conference-participant charge.

Billing increments matter too. Twilio's voice billing rules round each Programmable Voice call up to the next whole minute. If a flow uses two Call SIDs, Twilio's identifiers for individual calls, each is billed and rounded separately. This is not a universal carrier rule. It explains why many 20-second calls can cost more than one long call with the same aggregate duration.

Speech, model, and runtime

A cascaded voice agent usually has three AI meters:

• STT converts caller audio to text and is commonly priced by processed audio time. Channel count can change the quantity submitted.
• The LLM reads instructions, conversation history, retrieved knowledge, and tool results, then produces response tokens.
• TTS turns the response into audio and may be priced by characters, tokens, or generated audio time.

The runtime coordinates turn-taking, interruptions, tools, routing, and call state. Some platforms bundle speech into the runtime rate. Others bill provider usage separately. A native speech-to-speech model uses its own audio pricing, so separate STT and TTS belong in the budget only when the architecture actually invokes them.

Operations and fixed requirements

Production cost continues after the call. Include phone numbers, concurrency, recordings, storage, analytics, automated evaluation, test calls, support, security requirements, and data retention. Then include the people who maintain integrations, review failures, update knowledge, test releases, monitor completion, and handle incidents.

Retell AI pricing: separate the configuration from the carrier

Retell's voice infrastructure costs $0.055 per minute. Retell Platform Voices add $0.015 per minute, making those two components $0.07 per minute before the LLM, carrier, and optional features. Premium voice and model selections can exceed the $0.07–$0.31 headline range.

Retell-provided US telephony is listed separately at $0.015 per minute. With custom telephony, Retell adds no telephony charge, but your carrier still bills its side of the call. The AI-agent fee stops after a completed transfer while telephony continues. Failed-to-connect calls have no Retell AI-layer fee; connected voicemail, silence, and hold time remain billable while the agent is active.

Capacity and usage need separate lines:

• Pay-as-you-go includes 20 concurrent calls per workspace. Additional standard capacity is $8 per slot per month.
• Optional concurrency burst adds $0.10 per minute to the entire call that starts above the standard limit.
• Batch Call adds $0.005 per dial. AI Quality Assurance costs $0.10 per analyzed minute after the first 100 analyzed minutes per workspace.

Normal call duration is tracked to the nearest second without a generic per-call round-up. Retell's billing exceptions apply a 10-second minimum specifically to calls shorter than 10 seconds with a dynamic AI-first opening. LLM context above 4,000 tokens can trigger proportional billed-duration scaling. That exception should not be treated as a universal multiplier across every line item.

Calculate your monthly AI voice agent cost

Use this formula as the starting point:

monthly total cost =
  connected minutes x loaded connected-minute rate
  + attempted-call, transfer, and event fees
  + fixed platform, number, capacity, support, and compliance fees
  + implementation, engineering, review, and handoff labor

cost per successful outcome = monthly total cost / successful outcomes

The loaded connected-minute rate combines costs that follow the same connected-minute clock. When clocks or units differ, calculate each component separately before adding it to the monthly total. Do not charge for speech twice when a platform already bundles it.

1. Measure attempts, connects, duration, and outcomes

For outbound calls, keep attempted, connected, transferred, and successful calls as separate fields. A carrier connection can reach a person, voicemail, or an automated menu. It is not necessarily a successful business outcome.

For inbound work, start with connected calls and their duration distribution. Average handle time alone cannot reproduce per-call minimums and rounding.

2. Convert each component to its own billable unit

Use provider usage records where available:

telephony cost = sum(billable minutes for each call leg x that leg's rate)
STT cost = processed audio minutes across billed channels x STT rate
TTS cost = billable characters or audio units x TTS rate
LLM cost = input tokens x input rate + output tokens x output rate
runtime cost = billable agent minutes x runtime rate

Before launch, representative transcripts provide token and character assumptions. Once traffic is live, API usage values replace those estimates. Model output quantities must include any billed reasoning tokens, not just the words spoken to the caller.

3. Add attempts, transfers, and optional features

Per-event charges can be material at high volume. Retell's batch-dial fee and Bland's outbound-attempt minimum use different units and triggers. A failed connection can avoid an AI-runtime fee while still creating a carrier or dialer charge.

Keep knowledge retrieval, denoising, guardrails, personal-data removal, quality assurance, recording, and messaging separate unless the chosen plan includes them.

4. Size peak capacity separately

Monthly minutes do not reveal how many calls arrive at the busiest moment. Track peak active calls and call-start rate in short intervals. Capacity limits can cause queuing, rejection, throttling, or higher-priced burst usage.

Vapi's package limits distinguish Usage only's 4 concurrent calls from Core's 10; extra lines cost $10 per month. Retell's included 20 slots apply per workspace. ElevenAgents uses plan-specific capacity and optional burst pricing. Our Growth plan has no per-line fee, which is different from a promise of unlimited immediately available capacity.

5. Add fixed and operating costs

List platform fees, phone numbers, commitments, support, compliance, storage, monitoring, and internal labor separately. A $1,000 fixed requirement adds $1 per minute at 1,000 monthly minutes and $0.01 per minute at 100,000.

Keep one-time implementation separate from recurring maintenance. Amortize it over an explicit period if you need a monthly planning figure.

6. Reconcile to outcomes

Join the billing export to call outcomes. Track cost per attempt, connected call, transfer, and successful result. Cost per successful outcome includes quality: retries, extra turns, and avoidable handoffs can make a lower minute rate more expensive for the same task.

Three monthly workload scenarios

These illustrative workloads use selected component prices and explicit planning assumptions. They are not Dasha prices, vendor quotes, observed performance, or complete monthly budgets.

AI-active duration means the total assumed time the agent participates, including any warm-handoff overlap. Human-only transfer time is separate. The overlap is counted in human labor below because both the person and agent are present; it is not added again to AI minutes.

For runtime, assume one billable Vapi hosting minute per AI-active minute under the current new-customer Usage-only plan, with 4 included concurrent calls and $10 per additional line per month. This planning proxy does not specify how a provider bills transfers or idle connected time.

For speech, assume one submitted Deepgram Flux English audio channel per AI-active minute at the $0.0065 streaming pay-as-you-go promotional rate. The listed regular rate is $0.0077. Aura-2 costs $0.03 per 1,000 input characters; assume 900 characters per AI-speaking minute. These direct-provider quantities do not establish Vapi's channel billing.

For the model, GPT-5 mini pricing is $0.25 per million uncached input tokens and $2 per million output tokens. Assume the output quantities above include all billed reasoning tokens. No cache discount is applied.

For labor, assume $92 per engineering hour, $29.50 per operations hour, and 20 minutes of manual review per selected call. These are planning rates, not current wage estimates or vendor charges.

These components do not form a complete monthly budget. Carrier calls and transfer legs, answering-machine detection, numbers, conference participants, recording, storage, logs, evaluation, support, compliance, taxes, and implementation remain outside these amounts. Those quantities depend on call topology, duration distributions, retention, and the selected agreement. Included credits and free allowances are also omitted.

Reproduce the growing-production example

connected calls = 20,000 x 45% = 9,000
AI-active minutes = 9,000 x 4 = 36,000
runtime = 36,000 x $0.05 = $1,800
extra capacity = (11 - 4) x $10 = $70
STT = 36,000 x $0.0065 = $234
TTS = 36,000 x 50% x 900 / 1,000 x $0.03 = $486
LLM input = 36,000 x 3,000 / 1,000,000 x $0.25 = $27
LLM output = 36,000 x 200 / 1,000,000 x $2 = $14.40
engineering = 60 x $92 = $5,520
manual review = 9,000 x 3% x 20 / 60 x $29.50 = $2,655
handoff labor = 9,000 x 12% x (5 + 0.75) / 60 x $29.50 = $3,053.25

Core provides a separate capacity sensitivity: $29/month plus usage, with 10 included calls. At planned concurrency of 1, 11, and 73, package plus extra-line fees are $29, $39, and $659. Usage only's corresponding capacity fees are $0, $70, and $690. Neither option's capacity charge includes runtime or provider usage.

Telephony needs its own calculation. At Twilio's US outbound rate of $0.014 per Call SID minute before volume or negotiated discounts, use each leg's rounded duration, including any transfer continuation. Answering-machine detection is $0.0075 per completed call with the feature enabled, not per attempted dial. Conference-participant minutes are another meter. Multiplying aggregate AI minutes by the outbound rate would miss those boundaries.

For outcome sensitivity, assume 50% of the growing scenario's 9,000 connected calls succeed: 4,500 outcomes. Each $1,000 of fully reconciled monthly cost then contributes about $0.22 per outcome. This assumption is not a success-rate benchmark, and dividing the partial component table by 4,500 would not produce total cost per outcome.

Every $0.01 change in a rate applied to all AI-active minutes changes modeled spend by $21, $360, and $3,000 across the three workloads. One extra conversational minute can affect several layers at once.

Seven costs that change the invoice

1. Failed calls, voicemail, and short conversations

Each layer can start billing at a different event: dial, ring, carrier answer, or agent activation. Our agent layer does not charge for attempts that fail to connect. That does not waive carrier charges. Retell's AI-layer exemption likewise differs from Bland telephony's outbound-attempt minimum.

2. Minimums, rounding, silence, and hold time

Apply minimums to the component that defines them. Telnyx's 60-second voice-engine increment is separate from carrier and model billing. Retell's dynamic-opening minimum is conditional. ElevenAgents' silence discount applies only to qualifying silence periods longer than 10 seconds; other layers can continue billing.

3. Transfers and overlapping call legs

A transfer may use a carrier handoff, a bridge, a conference, or a new outbound call. The AI can leave while phone legs and human labor continue. A warm handoff can temporarily involve three participants. Price the actual topology rather than assigning one universal transfer multiplier.

4. Model, prompt, and voice choices

Long prompts, replayed history, retrieved knowledge, and tool output increase input usage. Premium voices or models need to justify their additional cost through the outcome they help deliver. Cached-input pricing applies only where the provider and request qualify.

5. Concurrency, throughput, and burst behavior

Peak active calls and call starts per second are different limits. Queueing, failover headroom, test traffic, and burst behavior can change the capacity budget even when monthly minutes stay constant.

6. Recording, storage, analytics, and evaluation

Retention accumulates as each month's recordings are added. Evaluation can have a separate per-analyzed-minute rate. Log-ingestion and indexing prices do not cover application-performance traces unless that service is also included. Budget coverage and retention from operating needs rather than assuming every call must use every add-on.

7. Support, compliance, integration, and maintenance

Dedicated support, service-level agreements, single sign-on, data residency, custom retention, and launch help may require a contract or fixed add-on. A custom stack must fund equivalent integration, monitoring, release, security, and on-call responsibilities.

Compare bundled, modular, and do-it-yourself pricing

Our managed voice AI backend is for technical teams that want a managed runtime without operating every component themselves. A fully custom stack can be the right choice when owning those components is a product requirement; our build-versus-buy breakdown separates those responsibilities.

When one supplier provides carrier, speech, model, and agent layers, a single invoice can simplify procurement. Switching still involves the relevant phone-number routes, speech and model configuration, prompts, tools, and stored operational data. Compare ownership and migration work alongside the usage rate.

None of these models is automatically cheapest. Scale, provider requirements, engineering capacity, reliability targets, and operational ownership determine the tradeoff.

How Dasha pricing fits the model

We help technical teams build and run production voice AI agents through a managed runtime, APIs, and a web application. Our pricing separates the agent layer from third-party carrier and model usage:

• Growth starts at $0.08 per minute, excluding VoIP and LLM tokens.
• We bill agent usage in one-second increments with no minimum call duration.
• Attempts that fail to connect do not create an agent-layer charge.
• Growth has no per-line fee, with volume and committed-usage discounts available.

A Dasha forecast adds the required carrier, model, optional-service, and internal operating costs. “Connected” billing is a commercial trigger, not proof that the agent completed its business task. Keep business outcomes as a separate denominator.

Use this quote-comparison checklist

Ask every shortlisted provider the same questions:

1. What does the published rate include: runtime, STT, TTS, LLM, telephony, numbers, recording, and support?
2. When does billing start and stop for an attempt, connected call, voicemail, silence, hold, and transfer?
3. What is the minimum or rounding increment for each component and each call leg?
4. How do destination country, direction, local versus toll-free numbers, and carrier surcharges change the rate?
5. How do model, prompt length, conversation history, voice, language, and optional features change spend?
6. What concurrency and call-start rate are included, and what happens at the limit?
7. Which recording, retention, analytics, evaluation, privacy, and compliance requirements add fixed or usage fees?
8. Which support, uptime, dedicated-capacity, or implementation terms require an enterprise agreement?
9. What commitments, renewal terms, overages, ramp periods, and unused-volume rules apply?
10. Can usage exports be joined to business outcomes so you can measure cost per successful result?

Build the first estimate from unit prices, then replace planning assumptions with pilot traffic and agreed commercial terms. That produces a defensible budget without treating a starting minute rate as the cost of the whole workflow.

Start a Dasha technical evaluation to measure your workload's duration and completion patterns. Our Developer plan includes 1,000 free minutes and one concurrent call, with no credit card required. Keep carrier and model usage separate in the evaluation budget.