What Voice AI Actually Costs at Scale: Per-Minute Pricing, Hidden Cost Lines and Total Cost of Ownership

AI voice agent cost in production is the sum of telephony, speech-to-text, model inference, text-to-speech, and platform orchestration, plus number rental, failed-call billing, retries, concurrency, storage, and engineering time. At 50,000 minutes a month, the published rates in this article put usage between about $2,075 and $8,250 before fixed fees. Headline per-minute rates rarely include every line. Multiply your monthly minutes by each published component rate, then add fixed fees, to estimate total cost of ownership.

Teams past pilot often fixate on one quoted rate. That number is only a starting point. Once real traffic arrives, origination and termination, speech pipelines, model choice, and platform margin stack differently by volume, inbound versus outbound mix, and whether the vendor owns the telephony leg. Fixed lines such as number rental and minimum commitments can dominate at low volume. Retry logic can multiply outbound minutes far beyond the contact list size.

This article decomposes what a per-minute rate covers, compares published models for Plivo, Vapi, Retell AI, Bland AI, and ElevenLabs, works three volume scenarios with copyable arithmetic, and lists the cost lines pricing pages omit. It closes with a method for modeling return against your own baseline, not a vendor claim.

What a Per-Minute Rate Actually Covers

A quoted per-minute rate is almost never a single service. It is a stack. Telephony origination or termination moves audio on the public network. Speech-to-text (STT) turns caller audio into text. A language model (LLM) decides what to say. Text-to-speech (TTS) renders the reply. An orchestration layer ties turn-taking, tools, and call control together and usually carries the platform’s margin.

Bundled pricing publishes one blended per-minute figure that is meant to cover most or all of that stack. Unbundled pricing meters components separately so you can swap STT, TTS, or model providers. Neither shape is automatically cheaper. Bundled rates simplify forecasts and hide internal markups. Unbundled rates expose which line drives spend and reward teams that negotiate upstream contracts, but they also make invoices harder to predict until traffic patterns stabilize.

Component prices show why the stack adds up quickly. In Vapi’s current public calculator, transcription runs from $0.0095–$0.0099/min, a language model from $0.0077–$0.0452/min depending on the model, and an ElevenLabs voice from $0.0146–$0.0238/min. Those are inputs, not a full call cost. You still add telephony, orchestration, and any platform fee.

Ownership of the telephony leg changes both cost and control. When a platform operates its own network, origination and termination can sit closer to the agent runtime and avoid a reseller markup on every minute. When a platform resells another carrier, that markup is embedded in the quote or appears as a separate line. Buyers should ask which leg the vendor owns, whether bring-your-own SIP trunking is allowed, and whether the per-minute rate changes if you supply the carrier path yourself. On Plivo pricing, the Voice AI agent rate is $0.03/min and covers voice models, transcription, and orchestration, with telephony billed separately. On Plivo’s voice pricing page, United States local calls are $0.0115/min outbound and $0.0055/min inbound, and SIP trunking starts lower, at $0.0046/min for US local outbound.

The Comparison Table

Use the tables as a map of pricing shape, not as a forever-accurate rate card. Every rate below was read from the vendor’s own public pricing page in October 2026. The entries reflect what each vendor publishes on its own site; Voice AI list prices change often, so confirm current rates with the vendor before you sign.

Pricing model and headline rate

What the rate covers

Published rate detail

• Plivo. Voice AI agent $0.03/min excluding telephony. United States Voice API calls $0.0115/min outbound and $0.0055/min inbound for local numbers, $0.0060/min outbound and $0.0180/min inbound for toll-free. Concurrent calls, the Vibe Agent builder, simulation testing, transfer to a human, and a knowledge base are listed as included at no extra charge. Pay-as-you-go includes 50 concurrent calls and $10 in free credits; Enterprise starts at $1,000/month.
• Vapi. Hosting $0.05/min, with STT, LLM, TTS, and transport billed at each provider’s listed price on top. Vapi’s own estimate for 1,000 minutes is $82–$129 in total, about $0.082–$0.129/min all-in. Packages run from usage-only ($0, 4 concurrent calls) to Core ($29/month, 10 concurrent calls) and Pro (10% of hosting, $999/month minimum). HIPAA is a $2,000/month add-on.
• Retell AI. Public range $0.07–$0.31/min depending on model, voice, and add-ons. The default calculator example is $0.15/min: $0.055 voice engine, $0.08 LLM, $0.015 TTS, $0.00 telephony. Retell-supplied US telephony adds $0.015/min and numbers are $2.00/month. 20 concurrent calls are included; extra capacity is $8 per concurrent call per month.
• Bland AI. Start $0.14/min with no platform fee; Build $0.12/min plus $299/month. The rate covers the LLM, STT, and TTS; telephony is billed separately on your own carrier or Bland’s at pass-through cost. Transfer minutes are $0.05/min on Start and $0.04/min on Build. Build caps usage at 50 concurrent calls and 2,000 calls per day.
• ElevenLabs. Plans from $6/month (Starter, 75 minutes) to $990/month (Business, 12,375 minutes), with additional minutes at $0.08/min on every plan and $0.16/min burst pricing above the concurrency limit. LLM usage is billed on top by model, and ElevenLabs adds no telephony fee: your own carrier bills you directly.

Read the “included” column as carefully as the rate. Vapi’s $0.05/min is a hosting fee, not an all-in production cost; Retell’s range widens with model and voice choice; Bland’s bundled minute sits beside a plan fee on Build and still excludes telephony; ElevenLabs’ $0.08/min excludes the model. Plivo’s structural edges for cost modeling are owned telephony and no model lock-in on the Plivo AI Agents platform.

Worked Cost Scenarios

The useful artifact is arithmetic you can copy. Pick a monthly minute volume, multiply each published line by that volume, then add fixed monthly fees. Every row below is minutes multiplied by the stated rate, and every row says which layers it covers, because a hosting-only fee and an all-in rate are not the same number.

How to run the worksheet

1. Forecast answered minutes (inbound) or connected talk time (outbound), not dials.
2. Multiply minutes by each usage line you will actually enable (platform fee, STT, LLM, TTS, telephony).
3. Add monthly platform fees, number rental, storage, and concurrency charges.
4. For outbound, multiply again by expected attempts per successful connect (your measured retry factor, not a vendor default).
5. Add engineering hours at your fully loaded internal rate.

The scenarios use US outbound calling. The rates used:

• Plivo: $0.03/min agent + $0.0115/min US local outbound = $0.0415/min all-in.
• Vapi: its published all-in estimate of $0.082–$0.129/min.
• Retell AI: its $0.15/min calculator default + $0.015/min Retell telephony = $0.165/min.
• Bland AI: $0.14/min on Start and $0.12/min on Build, both excluding telephony.
• ElevenLabs: $0.08/min, excluding LLM and carrier. Each ElevenLabs plan fee equals its included minutes at $0.08/min, so the effective rate is $0.08/min whichever plan you pick.

Scenario A: Small deployment (5,000 minutes per month)

At this tier, fixed fees and number rental can rival usage. Plivo lists local numbers at $0.50/month; Retell lists its numbers at $2.00/month. A small team that needs Vapi’s Core package for more than 4 concurrent calls adds $29/month, and a team on Bland’s Build tier adds $299/month before any minutes run. Bland and ElevenLabs totals rise further once telephony, and for ElevenLabs the LLM, are added.

Scenario B: Mid-sized deployment (50,000 minutes per month)

Vapi’s hosting fee alone is $2,500 at this volume ($0.05 × 50,000), which is why its all-in estimate matters more than the headline. Retell’s public range of $0.07–$0.31/min means the same 50,000 minutes spans $3,500–$15,500 depending on where your model and voice choices sit inside that band. For ElevenLabs, 50,000 minutes on the Business plan is $990 for the first 12,375 minutes plus 37,625 additional minutes at $0.08, or $4,000 before the LLM and carrier.

Scenario C: High-volume outbound (200,000 minutes per month)

At this volume, concurrency and plan caps start to decide the vendor as much as the rate. Plivo’s pay-as-you-go plan covers up to $2,500 of monthly usage, so this volume runs on its Enterprise plan (from $1,000/month, custom pricing and volume discounts). Bland’s Build tier caps usage at 2,000 calls per day, so the 200,000-minute scenario exceeds that cap when average connected call length is under about 3.3 minutes. The Build line is therefore a reference point for shorter-call programs, which need a higher Bland tier (Scale, listed in Bland’s billing docs, or custom Enterprise pricing). Vapi’s Pro package charges 10% of hosting with a $999/month minimum, which adds $1,000/month here. Retell charges $8 per concurrent call per month above the 20 included.

Outbound changes the math more than inbound. If each successful conversation requires multiple dial attempts, and failed or short connects still bill, your contact list size is the wrong denominator. Track average attempts to connect in your dialer, then multiply expected talk minutes by that factor. Billing rules for ringing, voicemail, and early hangups differ by vendor, so ask for the exact start-of-billing event in writing.

Outbound programs may also need consent, calling-hour, recordkeeping, and call-abandonment controls. Check the FTC Telemarketing Sales Rule and the FCC ruling applying the TCPA to AI-generated voices, then have counsel confirm which requirements apply to your use case.

Substitute your volumes freely. The pattern stays the same: published unit rate times minutes, plus fixed fees, plus retry factor, plus people cost.

The Charges a Pricing Page Leaves Out

Number rental scales with caller IDs and geography, not average minutes. Plivo lists US local numbers at $0.50/month and toll-free numbers at $1.00/month on voice pricing; Retell lists its numbers at $2.00/month. US numbering also carries administration costs that carriers fund under 47 CFR Part 52, and providers may charge to port a number (FCC number-porting guidance).

Failed, abandoned, and voicemail-answered calls may still bill. Model attempts separately from connected conversations. Concurrency can also add cost: Vapi includes 4 calls on usage-only and 10 on Core, then charges $10 per extra line monthly; Retell includes 20 and charges $8 per extra call monthly; Bland caps Start at 10 and Build at 50. ElevenLabs charges $0.16/min above the plan’s concurrency limit.

Recording retention grows with volume. Plivo lists recording at $0.0000/min, storage free for 90 days, then $0.0004/min monthly. Production teams should also budget governance work against frameworks such as the voluntary NIST AI RMF.

Engineering can outweigh the minute bill. Integration, evaluation, monitoring, and incident response continue after launch. The U.S. Bureau of Labor Statistics reports a May 2025 median annual wage of $135,980 for software developers (BLS software developers). Use the build versus buy framework to test whether full-code, low-code, or managed delivery fits that people cost.

Cost at Low Volume

Below roughly ten thousand minutes per month, variable per-minute charges often stop being the main story. Minimum commitments, platform subscription fees, included-minute bundles, and number rental take over. Bland’s Build tier attaches a $299/month platform fee before a single minute runs. ElevenLabs Agents pairs plan fees (from $6/month on Starter to $990/month on Business) with included minutes and $0.08/min additional minutes. Vapi lists HIPAA-eligible data handling as a $2,000/month add-on, and its Pro package carries a $999/month minimum. Those fixed lines can exceed usage for a clinic, agency, or regional support desk that only needs a few thousand minutes.

What makes a platform affordable at low volume is usually structural, not the smallest headline rate:

• True pay-as-you-go with no monthly minimum
• No mandatory platform subscription before production features become available
• Cheap or included local numbers for the concurrency you actually need
• Clear rules on whether short calls, failed dials, and tests bill
• Room to grow into volume discounts without a punitive recontract

Plivo’s public pricing is pay-as-you-go with $10 in free credits and 50 concurrent calls, and an Enterprise tier that starts at $1,000/month for coverage across 190+ countries, custom pricing and concurrency, and compliance needs such as BAA and HIPAA. Match the tier to your volume instead of buying enterprise fixed cost for a pilot. If you are still proving containment, keep the contract shape flexible and instrument costs per resolved conversation, not only cost per minute.

Modeling Return Honestly

Return modeling needs three measured inputs: fully loaded cost per human-handled call, automatable volume, and production containment. BLS lists a May 2025 median hourly wage of $21.53 for customer service representatives (BLS customer service representatives). Its June 2026 compensation release puts private-industry employer cost at $46.89 per hour, with benefits at 30.0% (BLS ECEC). Use your own geography, benefits, occupancy, and handle time to convert hourly cost into cost per call.

A practical sequence:

1. Baseline cost per human-handled contact (wage, benefits, tooling, occupancy).
2. Identify intents safe to automate and the minutes those intents consume.
3. Apply a containment rate you measure in a controlled pilot, not a brochure figure.
4. Subtract AI variable cost, fixed platform fees, numbers, storage, and allocated engineering.
5. Re-run the model monthly as prompts and caller behavior change.

Do not import ROI multiples from marketing pages. Use the Voice AI pilot playbook to gather containment and cost-per-contact evidence, then apply the platform evaluation guide when vendors look similar.

Frequently Asked Questions

What does a first deployment actually cost?

Budget usage minutes times each published component rate, plus number rental, storage, fixed platform fees, and engineering time. At 5,000 minutes a month, published all-in rates put usage between roughly $200 and $825 before fixed fees, depending on the platform.

And what does that become at volume?

At 200,000 minutes a month, the same published rates range from about $8,300 to $33,000 in usage alone. Retries, concurrency charges, and plan caps then decide much of the remaining cost for outbound.

Bundled or unbundled: how do the models differ?

Bundled models publish one rate meant to cover most of the stack. Unbundled models charge a platform fee and pass through STT, LLM, TTS, and telephony. Hybrid plans add monthly fees on top of per-minute usage.

Which option suits a small, low-volume team?

Prefer pay-as-you-go without monthly minimums, low number rental, and clear short-call billing. Fixed platform subscriptions often cost more than minutes when volume stays low.

Where does the return come from against manual dialing?

Compute your fully loaded cost per human contact, multiply automatable volume by measured containment, and subtract the full AI stack including retries. Use your baseline wages and handle times, not vendor ROI claims.

Do failed calls and voicemail attempts count toward AI voice agent pricing?

Often yes. Some vendors bill from call progress or impose a per-attempt minimum, such as $0.015 per outbound attempt on some bundled telephony. Confirm the start-of-billing event and voicemail behavior in the contract before outbound launch.

What should I ask vendors before signing a Voice AI pricing contract?

Ask what the rate bundles, what is metered separately, failed-call and retry billing, concurrency caps or fees, number rental, storage retention, minimum commitments, and whether bring-your-own telephony changes the rate.

Conclusion

Treat the quoted minute rate as one row in total cost of ownership. Compare current, like-for-like rates; add numbers, storage, concurrency, failed-call rules, retries, and engineering; then anchor return math in measured cost per contact and containment.

Review Plivo pricing and the AI Agents platform, then validate fit with a structured pilot. You can model costs at signup or request a trial against real traffic.