What an AI Voice Agent Costs to Run at Volume
Last updated on July 3, 2026
We build and advise on production AI systems. Bring your questions to a free intro call.
Talk to usThe advertised “from $0.05/min” is one metered clock of several. A production voice agent runs closer to $0.12–$0.33 per connected minute once speech recognition, the language model, speech synthesis, and telephony land on top of the platform fee. The headline rate covers orchestration alone. The other four layers bill separately, and most meter every second the call is connected, silence included.
A forecast built on minutes × advertised rate under-budgets by two to three times. To get a number you can plan against, model the bill from your own call-length distribution and peak concurrency, then divide by resolved calls rather than raw minutes. Every rate in the model traces to a current 2026 source.
What the monthly bill is actually made of
The per-minute rate is one term of eight. A production voice agent bill looks like this:
monthly bill =
connected minutes × all-in runtime rate
+ attempt minimums (failed / short calls)
+ concurrency slots (reserved capacity)
+ phone numbers
+ transfer minutes (human handoff)
+ recording + storage + transcription
+ evaluation minutes (QA / judging)
+ compliance fixed feesThe all-in runtime rate itself is the sum of five metered layers. Each has its own unit and its own vendor.
| Layer | Representative 2026 rate (primary source) |
|---|---|
| Speech-to-text | Deepgram Nova-3 Monolingual $0.0048/min; Flux English $0.0065/min; AssemblyAI Universal-Streaming $0.15/hr |
| Language model | ~$0.01–$0.03/min on a mini-tier model, ~$0.05–$0.08/min on a flagship (token-derived) |
| Text-to-speech | Deepgram Aura-2 $0.030 / 1k characters; ElevenLabs $0.10 / 1k characters, $0.05 on Flash/Turbo |
| Telephony | Twilio local $0.0140/min outbound, $0.0085/min inbound; SIP $0.0040/min |
| Platform / orchestration | Vapi $0.05/min; Retell voice infra $0.055/min; Bland $0.11–$0.14/min; LiveKit $0.0100/min; Pipecat $0.01/min |
Two units need translating before they compare. TTS is priced per character. A voice agent speaking at roughly 150 words per minute produces about 800–900 characters of speech per minute of its own talk time, and the agent speaks for roughly half the call. Aura-2 at $0.030 / 1k characters therefore costs about $0.024–$0.027 per minute of agent speech.
The language model is priced per token, and a voice turn resends the whole conversation each time, so per-minute LLM cost is dominated by input tokens (the system prompt plus growing history) multiplied by turns per minute. A 5,000-token system prompt with a few thousand tokens of dialog, around two turns per minute, uncached, runs about $0.01–$0.03/min on a mini model and $0.05–$0.08/min on a flagship. Prompt caching collapses the system-prompt share, which is why it is the single largest LLM lever.
Three stacks, three volumes: what you will pay
The table below models three stack archetypes across three volume tiers. Assumptions: US phone calls, no SMS, no HIPAA add-on, 100% connected minutes, no committed-use discounts. “Economical” uses fast low-cost models; “premium” uses a flagship LLM or a premium voice. These are our estimates, built up from the published per-unit rates above, not vendor quotes.
| Monthly connected minutes | Managed platform (economical) | Managed platform (premium) | Platform + bring-your-own models | Self-assembled cloud |
|---|---|---|---|---|
| 1,000 | $110–$140 | $170–$310 | $75–$110 | $35–$80 |
| 10,000 | $1,100–$1,400 | $1,700–$3,100 | $750–$1,100 | $350–$800 |
| 50,000 | $5,500–$7,000 | $8,500–$15,500 | $3,750–$5,500 | $1,750–$4,000 |
The self-assembled column looks cheapest, and on usage alone it is. It omits engineering labor, on-call ownership, incident response, and compliance evidence. That gap is the whole breakeven question.
One thing the table hides: the fixed items sit on top of every column. At 1,000 minutes a month, usage at $0.15/min is $150, but a single concurrency floor, a HIPAA add-on, or a committed-tier seat can each exceed the entire usage bill. Below roughly 5,000 minutes a month the bill is fixed-fee-dominated. Stay on pay-as-you-go, skip committed and business tiers, and optimize the fixed floors rather than chasing a lower per-minute rate. The per-minute rate is not what you are paying for yet.
Two worked examples: the distribution is the story
Budgeting on the average minute misleads, because cost is right-tailed and the shape of your call mix moves the bill more than the rate does. Two examples at the same 10,000-minute scale, on the same platform-plus-BYO stack, produce very different invoices.
Inbound support, 10,000 connected minutes, 4-minute average call
That is about 2,500 calls. The all-in runtime rate builds up as:
| Line item | Rate | Monthly |
|---|---|---|
| Platform (Vapi) | $0.05/min | $500 |
| STT (Deepgram Nova-3) | $0.0048/min | $48 |
| LLM (mini-tier) | ~$0.02/min | $200 |
| TTS (Deepgram Aura-2) | ~$0.025/min | $250 |
| Telephony (Twilio inbound) | $0.0085/min | $85 |
| Concurrency (5 lines above the 10 included) | $10/line | $50 |
| Phone number | $1.15/mo | $1 |
| Evaluation (15% sample, judged) | ~$0.10/min | $150 |
| Total | ~$1,285 |
That is about $0.13 per minute all-in, roughly double the $0.05 platform headline. Swap Aura-2 for a premium voice and the mini model for a flagship, and usage alone climbs past $0.25/min.
Outbound reminders, 10,000 dials, 88% voicemail
Here the workload is short and telephony rounding dominates. Assume 88% of dials hit voicemail (8 seconds each) and 12% connect (3 minutes each). Twilio rounds every call leg up to the next full minute, so 8,800 voicemail dials bill as 1 minute each and 1,200 connects bill as 3 minutes each: 12,400 billed minutes against roughly 4,780 actual. The AI stack runs only on the 3,600 connected minutes.
| Line item | Basis | Monthly |
|---|---|---|
| Telephony (Twilio outbound) | $0.0140 × 12,400 billed min | $174 |
| Answering-machine detection | $0.0075 × 10,000 dials | $75 |
| AI stack | ~$0.11/min × 3,600 connected min | $396 |
| Total | ~$645 |
Nearly a third of that bill goes to dials that never became a conversation: voicemail minutes rounded up to a full minute each, plus machine detection at $0.0075 a dial. The single highest-leverage fix is per-second telephony. Bland bills connected time “prorated to the exact second,” and Deepgram bills STT by the second. On short-call-heavy outbound, that alone removes most of the rounding tax. On multi-minute inbound calls, rounding is negligible and not worth engineering around.
The lesson generalizes: model the p90 call, not the mean. A 10-minute support call on a flagship model with several knowledge-base lookups can cost five to ten times a 2-minute booking on a mini model, and both sit on the same flat quoted rate.
Why the advertised rate doubles in production
Five mechanisms, in rough order of how much they surprise people.
The headline is one layer of five. Vapi’s pricing page states its $0.05/min “Excludes Model Provider Costs” and that “Model costs are passed on to customer.” The advertised number is the orchestration fee. STT, LLM, TTS, and telephony are yours to add.
Telephony rounds short calls up. Twilio’s changelog documents rounding “the duration of any completed sub-second call up to 1 second.” A reproducible API test against the same billing shows a 69-second call “being rounded up to minutes,” because Twilio “charges in full minute increments.” Ring time is not charged, but a voicemail pickup is, because the platform sets the call to in-progress and cannot distinguish a machine from a human answer. Every voicemail is therefore at least one billed minute, rounded up. This bites only when calls are short and numerous, as the outbound example showed.
Silence, hold, and ringing meter as talk time, and not only on the phone bill. Streaming STT and voice-agent bundles bill for connection duration. Deepgram’s Voice Agent API rates are “calculated based on websocket connection time” on every tier. So thirty seconds of dead air on a two-minute call is paid at the telephony layer, the STT layer, and the platform layer at once. Endpointing and turn-detection tuning cut all three simultaneously, which is why that knob compounds.
Failed and short calls carry minimums. Bland charges an “Outbound Minimum … $0.015 / call” and bills “Voicemail … as part of standard call time.” Every dial that never becomes a conversation still costs a floor. Outbound campaigns pay this on the whole no-answer tail.
Add-ons scale per minute, not once. They grow at the transcription or call rate, not as a flat fee. Retell lists Knowledge Base at ”+$0.005/minute,” PII Removal at “+0.01/min,” and AI Quality Assurance at “$0.10/min.” Deepgram redaction adds “$0.0020 per minute” on top of the base STT line. Retell also documents that agents “that use more than 4,000 LLM tokens in their prompts” have their billed duration scaled up, so prompt bloat is a direct cost problem, not only a latency one.
None of this makes doubling a law. It follows whenever the headline excludes components, which most developer-first headlines do. A platform that bundles telephony and models into one quoted rate will show a smaller gap.
Cost per resolved call is the number that runs your business
Per minute is the input. The unit the business actually runs on is cost per outcome: a resolved ticket, a booked meeting, a completed intake. The formula:
cost per resolved call = (all-in cost per minute × average handle time) ÷ resolution rateA $0.15/min agent on 4-minute calls is $0.60 per call. At a 55% resolution rate, the cost per resolved call is $0.60 ÷ 0.55 ≈ $1.09. Every failed or escalated call still consumed its minutes but resolved nothing, so it loads onto the calls that did resolve.
Containment and resolution measure different things. A call is contained when it never reaches a human, which includes the customer who gave up and hung up. It is resolved only when the problem is fixed. Vendors quote the higher containment number. In our production work, blended resolution across a typical inbound mix lands in the 45–65% range; high-structure call types like order status and appointment booking exceed 70–80%. Model the resolution rate you can actually verify, not the containment rate a dashboard reports.
This is why per-minute knob-chasing misses. Shaving $0.02/min off TTS while resolution sits at 45% optimizes the wrong variable. Lifting resolution from 45% to 60% cuts the cost per resolved call by a quarter with no change to the per-minute rate at all. For the make-or-buy comparison, a fully loaded US support agent runs roughly $3,000–$4,000 a month; the agent earns its keep only on the share of calls it resolves without a human.
The handoff keeps billing
Human transfer is a running cost, and it is easy to double-bill. A warm transfer opens a second call leg: the agent briefs the human on a new leg, then the two calls merge. That second leg bills. Bland’s docs state “Transfer time is billed additionally when the call is transferred to a human” and that merged calls are “Billed for transfer duration,” though transfer is “Free when using your own Twilio number.” So bring-your-own-carrier zeroes the platform’s transfer surcharge but not the underlying carrier minutes.
Model escalation as:
escalation cost = escalation rate × (second telephony leg + any continued AI / platform metering during the merge + the human agent's minutes)A 40% escalation rate means 40% of calls carry a transfer surcharge and consume a human agent. The agent’s entire saving lives in the share it resolves without one.
Five knobs that cut cost without degrading the agent
Ranked by leverage.
- Prompt caching. A voice agent resends the same system prompt every turn. Cached input is billed at “0.1x base input price” on Anthropic, and OpenAI lists cached input for its flagship at $0.25 / 1M tokens against $2.50 standard, a 10x discount. Google’s context cache discount is “90%” for Gemini 2.5 and later. Structure the prompt so stable policy hits cache and only per-turn state stays dynamic. Near-zero downside.
- Model routing by turn type. Route routine turns (classification, slot filling, confirmations) to a fast cheap model and reserve a flagship for turns that need reasoning; run heavy reasoning after the call, where latency does not matter. This cuts LLM spend several-fold without a caller-perceptible downgrade, provided routing errors are measured. It buys nothing when the domain is too narrow to justify multiple model paths.
- Shorter TTS output. TTS is often the largest runtime layer and is character-priced. Terser responses cut synthesized characters, LLM output tokens, and latency together. Flash and Turbo voices are half the price of premium multilingual ones: ElevenLabs lists “$0.05 (Flash/Turbo)” against $0.10 per 1,000 characters. Over a phone codec, the audible quality gap often disappears, so test premium voices on the actual PSTN path before paying for them.
- Endpointing and turn-detection tuning. This cuts silence billed as talk time across telephony, STT, and platform at once, and reduces re-transcription from false endpoints. See our voice agent latency budget guide for the mechanics.
- Per-second telephony or bring-your-own-carrier. For short-call-heavy workloads, moving off per-minute-rounded carriers removes the rounding tax the outbound example exposed. Irrelevant for multi-minute inbound.
Knobs that quietly buy little: chasing the cheapest STT when STT is already the smallest layer at well under a cent a minute; shaving LLM token price while ignoring TTS, which is often three to five times larger; and self-hosting a model at low GPU utilization, where idle hardware loses to usage-based API pricing. Our LLM selection guide covers the self-host economics in detail.
What scales non-linearly
Three costs break the minutes × rate model entirely.
Concurrency is sold per slot, and peak is not average. Monthly minutes tell you nothing about peak demand. Size it from the busy hour:
required concurrency = (calls started in the busy hour × average handle time in minutes ÷ 60) × headroom factor50,000 minutes a month spread across 24/7 averages about 1.16 concurrent calls. The same 50,000 minutes compressed into weekday four-hour outbound windows averages about 9.5 concurrent, before any burst safety. Slots are a fixed monthly cost independent of minutes used: Vapi includes “10 … + $10 / line / mo,” Retell includes 20 then charges “$8.00/Concurrency/month.” Provision for peak, pay for peak, use average.
There is a spend-control decision hidden in the cap. When you hit it, a platform either protects your budget by rejecting calls or protects call completion by charging a burst surcharge. Retell queues an inbound call and, “if no slot opens after about 40 seconds,” drops it; it also offers burst capacity at a “$0.10/min surcharge applied to the entire call duration.” Bland’s lower tiers reject over the cap so that “calls get rejected rather than incurring unexpected charges.” Decide which hurts more, a missed call or a higher per-minute rate, and set the cap accordingly.
Evaluation is a recurring cost, not a one-time setup. Running an LLM judge over calls is a second model pass per evaluated call. Retell prices this directly at “$0.10/min” for AI Quality Assurance. At 100% coverage that can equal or exceed the base agent cost. Sample instead: 5–20% of production calls, plus regression suites only on changed flows. Do not run the most expensive judge path on every minute by default.
Compliance is a step-change in the whole cost structure. It is not a percentage surcharge. Vapi publishes fixed add-ons, “HIPAA at $2K/month” and “Zero Data Retention at $1K/month.” Others gate it entirely: ElevenLabs states “BAAs are only available to enterprise customers.” A HIPAA deployment does not just add a monthly line; it forces every component onto an enterprise contract, which raises the per-minute component rates underneath it too. A 50,000-minute HIPAA support deployment that would run $10,000 in usage lands closer to $13,000–$16,000 all-in once the add-on, enterprise component pricing, extra concurrency, and retained recordings stack up.
Recording storage compounds separately. Twilio prices recording at “$0.0025 / min” but storage at “$0.0005 / min per mo,” a stock that grows every month you retain audio. Compliance-mandated retention over years means storage climbs even when call volume is flat.
Platform versus self-assembled: the honest breakeven
The per-minute delta between a managed platform and a self-assembled stack is the platform fee, roughly $0.04–$0.10/min after accounting for the open-source orchestration hosting a self-assembled stack still pays ($0.01–$0.03/min on LiveKit or Pipecat). At 50,000 minutes a month that is about $3,000 saved.
Set that against what self-assembly adds: building and maintaining provider fallback, telephony hardening, monitoring, evaluation, and on-call. A $3,000 monthly saving does not pay for even a fractional senior engineer to own that. The breakeven is set by headcount, not minute volume. A self-assembled stack only pays for itself once the monthly platform-fee saving exceeds the fully loaded cost of the person maintaining it, which is past roughly 100,000 minutes a month and often well above that.
So the self-assembled case is really about control and compliance: data residency, custom telephony, model routing you cannot get on a platform. Choose it for those reasons, not to save on the per-minute rate. For a vendor-by-vendor pricing comparison, see our voice agent platform guide. The one-time build cost that precedes any of this is a separate calculation, not part of the monthly run rate.
Instrument the pilot before you trust any forecast
None of the math above means anything without your own numbers. You cannot tell a cheap-per-minute vendor from a cheap-per-outcome system until you have measured the workload. Before scaling, log:
- attempted, connected, and completed calls, and successful outcomes
- average handle time and p95 handle time
- peak concurrent calls by hour
- voicemail and failed-call rate
- human transfer rate and transfer duration
- user speech minutes, agent speech minutes, and total connected minutes
- turns per call and LLM input/output tokens per turn
- prompt cache hit rate
- evaluation sample rate and recording retention duration
A one-week billing-instrumentation pass tells you more than any rate card. It gives you the call-length distribution, the peak concurrency, and the resolution rate that the bill actually depends on.
Run your own numbers through our voice agent cost calculator; it applies exactly this method to your volume and stack. When the answer needs judgment rather than arithmetic, which stack, which compliance posture, where the breakeven sits for your team, that is a cost-modeling conversation we have often. Talk to us and we will map it.
