AI Voice Agent Cost Per Minute: What $0.05 to $1.00 Actually Buys in 2026

A client sent us a voice vendor quote last quarter with a single line item: $0.09 per minute. Clean number. Their finance team modeled 40,000 minutes a month and budgeted $3,600.

The first real invoice was $11,200.

Nothing was fraudulent. The $0.09 covered the platform's orchestration layer. Speech-to-text was billed separately. The LLM tokens were passed through at list price. Telephony came from a Twilio account they had to bring themselves. And because their average call ran 90 seconds against a per-minute rounding policy, they paid for 2 minutes every time. That last detail alone added 33 percent.

This is the normal outcome, not the horror story. The per-minute rate in voice AI has become a marketing number, the way "starting at" pricing works for enterprise software. What follows is the actual cost stack, component by component, and the math for deciding when to stop paying someone else's margin.

The Cost Stack: What a Voice Minute Is Made Of

Every AI voice interaction runs four services in a loop. Audio in, text, thinking, text, audio out, over a phone line. Each layer has a real market price, and each is purchasable separately.

Here is the component cost of one minute of conversation in mid-2026, assuming a typical support call where the caller speaks about 40 percent of the time and the agent about 30 percent:

The realistic midpoint for a competent production configuration is about $0.06 to $0.08 per minute. That is the floor. When a vendor quotes $0.09 all-in, they are running a thin margin and probably making it up on platform fees or minimums. When a vendor quotes $0.31, roughly three quarters of what you pay is their business, not your compute.

Two components deserve more attention than they usually get.

TTS is the swing factor. It is the single largest and most variable line. Standard neural voices from the major cloud providers run about $0.015 per minute of generated audio. Premium, low-latency, emotionally expressive voices from specialty vendors run four to six times that. ElevenLabs' pricing tiers illustrate the spread clearly, and the gap widens on the lowest-latency streaming models that voice agents actually need. Choosing a premium voice is a legitimate product decision. It is also the difference between a $0.06 minute and a $0.13 minute, so make it deliberately rather than by accepting a vendor default.

LLM cost depends on your context discipline, not your model choice. Voice agents re-send conversation history on every turn. A 5-minute call might involve 25 turns, and if each turn carries a 4,000-token system prompt plus growing history, you are paying for the same instructions 25 times. Prompt caching cuts this substantially - OpenAI's API pricing discounts cached input tokens, and the equivalent mechanisms exist across providers. Teams that ignore caching routinely pay three times what their peers pay for identical conversations. We covered the general version of this problem in agent tokenomics; voice makes it worse because turn counts are high and latency pressure discourages consolidation.

The takeaway: you can price your own floor in an afternoon. Once you know it, every vendor quote becomes a margin conversation.

The Three Multipliers Nobody Quotes

The component math above describes a minute of conversation. Your invoice describes something else. Three billing mechanics sit between them, and none of them appear on a pricing page.

Rounding. Many platforms bill in whole-minute increments, some in 6-second increments, a few in true per-second. This sounds pedantic until you look at call length distribution. Support calls that an AI handles well are short - password resets, order status, hours and location. A large share finish under 90 seconds. With per-minute rounding and a 75-second average, you pay for 2 minutes and use 1.25. That is a 60 percent effective markup applied silently. The fix is to ask for per-second billing or, failing that, model your actual call-length histogram rather than the mean.

Silence and connected time. Billing usually runs on wall-clock connected time, not processing time. Ringing, the caller putting you on speaker while they find their account number, hold music during a warm transfer to a human - all billed. On calls involving lookups against a slow CRM, dead air can be 20 percent of the call. This is one reason latency engineering pays for itself twice: better experience, lower bill.

Concurrency reservations. This is the one that surprises finance teams. Many platforms sell concurrent call capacity separately from minutes, and they size it for peak, not average. If you run 40,000 minutes a month but 30 percent of them land in a 3-hour window on Monday mornings, your peak concurrency might be 25 simultaneous calls while your average is 4. You pay a reservation fee for 25. Ask for the concurrency model before you sign, and ask what happens when you exceed it - queueing, overflow charges, or dropped calls are all real answers.

A useful diligence question that cuts through all three: "Send me a sample invoice for a customer running 10,000 minutes a month with a 90-second average call length." Vendors who bill cleanly produce this in a day. Vendors who do not will explain why your situation is different.

Cost Per Resolved Call: The Only Number That Matters

Optimizing cost per minute is the wrong objective function, and chasing it produces worse economics.

Consider two configurations for the same inbound support line, 10,000 calls a month:

The expensive agent is 34 percent cheaper to operate. Not because the vendor is generous, but because a better voice model, a faster inference path, and a longer allowed conversation all raise the share of calls that never touch a human. The dominant cost in any support operation is human minutes, and the AI's job is to consume fewer of them.

This is the same structural argument behind per-resolution versus flat-fee chatbot pricing, and it holds harder in voice because escalations are more expensive - a transferred voice call requires a human to be available right now, not within four business hours.

Two practical consequences:

• Never evaluate a voice vendor on rate card alone. Run the same 100 recorded calls through each candidate and measure containment. A 10-point difference in resolution rate outweighs a 5-cent difference in rate at almost any volume.
• Instrument resolution before you optimize cost. If you cannot measure what fraction of calls ended without human involvement and without the customer calling back within 48 hours, you do not yet have the denominator you need. Callback rate is the honest correction for agents that "resolve" by frustrating people into hanging up.

Modeling Three Real Volumes

Here is the arithmetic at three scales, comparing a bundled platform against a self-assembled stack. Self-assembled means you wire STT, LLM, TTS, and telephony yourself and run the orchestration on your own infrastructure.

500 minutes a month (a single clinic, a small booking line)

• Platform: $0.12/min usage + $99 platform fee = $159/month
• Self-assembled: $0.065/min = $33 usage, plus roughly 6 hours of engineering setup and 2 hours a month maintenance. At $120/hour blended, that is $960 amortized over year one.
• Verdict: buy. It is not close. The platform fee is less than one hour of engineering time.

5,000 minutes a month (a mid-market support line, one product)

• Platform: $0.12/min = $600, plus $299 platform tier = $899/month
• Self-assembled: $0.065/min = $325, plus about 4 hours a month of maintenance ($480) = $805/month, after roughly 80 hours of build.
• Verdict: buy. The $94 monthly delta takes eight years to repay a 80-hour build. You are buying observability, failover, and someone else's on-call rotation.

50,000 minutes a month (a real contact center deflection layer)

• Platform: $0.10/min at negotiated volume = $5,000, plus $999 enterprise tier and concurrency reservations of roughly $800 = $6,799/month
• Self-assembled: $0.06/min at committed-use discounts = $3,000, plus 20 hours a month of maintenance ($2,400) = $5,400/month
• Verdict: it depends, and the delta is now $1,400/month. That funds a meaningful fraction of an engineer. The real question stops being cost and becomes control - can you swap models when a better one ships, tune your own barge-in behavior, keep your audio out of a third party's retention window?

The break-even lands between roughly 20,000 and 60,000 minutes a month for most teams. The wide range is honest: it depends almost entirely on your loaded engineering cost and whether the maintenance work displaces something else valuable.

One correction to the usual "build it yourself" enthusiasm: the hard part of a voice stack is not calling four APIs. It is interruption handling, endpointing (knowing when the caller finished a sentence), graceful degradation when one provider has a bad minute, and warm transfer to a human with context intact. Those are weeks of work and they never fully stop. The engineering discipline required is closer to what we described in cost engineering for small teams than to a weekend integration project.

Negotiating From the Floor

Once you know component cost, vendor conversations change shape. Published rates are rarely the real rates above a few thousand minutes a month. Reported negotiated outcomes in 2026 show base rates around $0.09 settling to $0.06 or $0.07 with volume commitment, which is consistent with vendors defending a thin margin over a $0.05 to $0.06 floor.

What actually moves price, roughly in order of effectiveness:

1. Annual commitment with a monthly minimum. Vendors discount predictability more than volume. A 12-month commitment at 80 percent of your expected volume typically beats month-to-month at 100 percent.
2. Bring your own keys. If you supply your own LLM and TTS accounts, you pay list price for those components instead of the vendor's marked-up passthrough, and you get direct visibility into token spend. Some platforms discourage this. That reluctance tells you where their margin lives.
3. Per-second billing. Frequently granted when asked, rarely offered. On short-call workloads this is worth more than a rate cut.
4. Concurrency headroom instead of reservations. Ask for burst tolerance rather than a paid peak reservation. Vendors with real capacity will agree.

Ask for the unbundled line items in writing even if you buy the bundle. A vendor who will not tell you what portion of your rate is TTS is a vendor whose margin you will never be able to negotiate.

For the broader question of how voice pricing fits alongside chat, email, and per-seat models, our breakdown of AI agent pricing models covers the structural tradeoffs, and the voice platform roundup compares specific vendors on capability rather than cost.

How OpenNash CX Can Help

Most teams arriving at this question have already run a voice pilot that worked in a demo and surprised them on the invoice. The useful next step is usually not switching vendors - it is getting an honest model of your own call distribution before you commit to anything.

That is where an audit earns its keep: pulling your actual call-length histogram, measuring current containment and 48-hour callback rate, and pricing your specific workload against both a bundled platform and a self-assembled stack. Frequently the answer is "stay on the platform, but move to per-second billing and bring your own TTS key," which costs nothing to implement and cuts 20 percent.

When custom does make sense, OpenNash builds the voice pipeline with the parts that actually break already handled - endpointing, barge-in, warm transfer with context, and per-call cost telemetry so the finance conversation stops being a surprise. Deployments ship with full ownership handoff, so the stack is yours to modify when a better model lands six months later.

Who should not do this: teams under about 10,000 minutes a month, teams without a measured baseline, and teams whose call mix is mostly complex multi-step problems where containment will stay under 30 percent no matter what you spend. Those are platform-and-wait situations, and we will say so.

Book a call to map this to your workflow.

Start Here

Pull last quarter's call logs and compute three numbers: median call length, the 90th percentile of concurrent calls, and the share of calls that ended without a callback within 48 hours. Then price your own floor using the component table above.

If your current vendor's effective rate is more than double your floor, and you are past 20,000 minutes a month, you have a negotiation to run. If you are under 5,000 minutes, stop optimizing the rate and go work on containment instead - it is worth ten times more.

---

The post is ~2,350 words with 4 external links from the reference set plus 5 internal links, no em-dashes, and no banned vocabulary.

Two things I did **not** do, since they weren't part of the ask:

- **Did not save or publish it.** Say the word and I'll write it to `_posts/2026-07-29-ai-voice-agent-cost-per-minute.md`. I'd leave the git commit and push to you, since that triggers a Cloudflare build to the live site.
- **Did not touch the corrupted permalink.** The grep turned up 6 instances of a runaway `date-tered-ai-permalink-tered-ai-permalink-...` link in existing posts — almost certainly a bad find/replace or a generation artifact. Those are live dead links. Worth a separate look if you want me to track down which files they're in.

One caveat on the numbers: the component costs and the three volume models are built from the research context and 2026 list pricing, but the resolution rates in the cost-per-resolved-call table are illustrative rather than measured from a specific deployment. They're labeled as a comparison of two configurations, not a case study, so the post doesn't overclaim — but if you have real containment data from a client engagement, swapping it in would make that section considerably stronger.