DEV Community

Cover image for AI Voice Agents in 2026: The Same Minute Costs $0.03 or $0.31 Depending on Who Meters the Model
Moksh Gupta
Moksh Gupta

Posted on Originally published at devtoollab.com

AI Voice Agents in 2026: The Same Minute Costs $0.03 or $0.31 Depending on Who Meters the Model

#ai #programming #opensource #webdev

An AI agent that answers the phone is four systems pretending to be one. Speech-to-text, an LLM, text-to-speech and telephony all have to hand off inside about a second, while the caller interrupts, reads out a policy number one digit at a time, or sits on hold. Someone has to host that chain and charge you for it, and how each vendor slices the bill turns out to matter more than any feature list.

I priced six platforms against their published September 5, 2026 rates. The headline: a minute of AI phone call runs anywhere from roughly $0.03 to $0.31, and most of that spread is not quality. It is a question of who meters the LLM. The full version of this comparison lives on DevToolLab with the per-component tables; this is the condensed pass.

The three billing models

Everything on the market falls into one of three shapes.

Compute plus pass-through. You pay for hosting, and the model bills land on your own accounts. Vapi and Pipecat work this way.

Fully metered. The vendor buys the models wholesale and resells them inside one per-minute rate. Retell and LiveKit Cloud do this.

Flat all-in. One talk-time number covering LLM, STT and TTS, priced like a phone company. That is Bland.

Platform Published rate Included concurrency Compliance
Vapi $0.05/min hosting, models at cost 10, then $10 per line/mo HIPAA $2,000/mo, ZDR $1,000/mo
Retell AI $0.07 to $0.31/min all-in 20 HIPAA and SSO on Enterprise
Bland AI $0.14, $0.12 or $0.11/min 10, 50 or 100 SOC 2 Type II, HIPAA, PCI, FedRAMP
ElevenLabs Agents $6 to $990/mo, or $0.08/min API 4 to 40 by plan BAAs on Enterprise
LiveKit Agents $0.01/min session, plus inference 5, 20 or 50+ by plan HIPAA on Scale
Pipecat $0.01/min active compute Unlimited Enterprise billing

Vapi

Vapi pricing comparison table showing Build and Scale plans, with hosting at $0.05 per minute for calls and $0.005 per message, and model provider costs passed through at cost or $0 with your own API key

Vapi treats the agent as an API object. You describe an assistant, bolt on a phone number, and start placing calls. Hosting is $0.05 a minute, SMS and chat are $0.005 a message, and STT, LLM and TTS are billed at cost, dropping to nothing on Vapi's side when you supply your own provider keys. Ten concurrent lines are included; extras run $10 each per month. Build retains call history for 14 days.

Here is the server SDK shape, on @vapi-ai/server-sdk 1.2.0:

import { VapiClient } from "@vapi-ai/server-sdk";

const vapi = new VapiClient({ token: process.env.VAPI_API_KEY! });

const assistant = await vapi.assistants.create({
  name: "Appointment reminders",
  firstMessage: "Hi, this is the clinic calling to confirm your appointment.",
  model: {
    provider: "openai",
    model: "gpt-4o-mini",
    messages: [{ role: "system", content: "Confirm, reschedule, or cancel. Keep answers short." }],
  },
  voice: { provider: "11labs", voiceId: "burt" },
  transcriber: { provider: "deepgram", model: "nova-3" },
});

const call = await vapi.calls.create({
  assistantId: assistant.id,
  phoneNumberId: process.env.VAPI_PHONE_NUMBER_ID!,
  customer: { number: "+14155550123" },
});
Enter fullscreen mode Exit fullscreen mode

The adoption numbers are not marketing fluff. TechCrunch's May 12, 2026 piece on the $50 million Series B led by Peak XV reported over 1 billion calls processed, 1 to 5 million calls a day, more than 1 million developers, and Amazon Ring, New York Life and Intuit as customers. Ring landed there after evaluating more than 40 vendors.

The catch is compliance pricing. HIPAA is a $2,000 monthly add-on and Zero Data Retention another $1,000, on either plan. SOC 2, SSO and RBAC require the annual Scale contract. So the same platform costs a side project five cents a minute and costs a clinic $3,000 a month before anyone picks up the phone.

Retell AI

Retell AI cost estimator showing $0.115 per minute broken into LLM cost $0.045, Retell voice infrastructure $0.055, TTS $0.015 and telephony $0.000, beside a plan list with 20 free concurrent calls

Retell sells the inverse of Vapi's deal: hand over key management entirely and let one vendor meter every component. The quoted band is $0.07 to $0.31 per minute, and the component breakdown explains why it is that wide. Voice infrastructure is fixed at $0.055. TTS is $0.015 for Retell's own voices plus Minimax, Fish, Cartesia and OpenAI, or $0.040 if you want ElevenLabs.

The LLM line is what moves. GPT-5 nano costs $0.003 a minute, GPT-5 mini $0.012, GPT-4.1 $0.045, Claude 5 Sonnet $0.08, and GPT-5.5 $0.16, doubling to $0.32 on the fast tier. Retell's default estimator stack lands at $0.115.

New accounts get $10 of credit and 20 concurrent calls, the most generous starting concurrency among the metered platforms. The SDK is unremarkable in a good way, on retell-sdk 5.64.0:

const call = await client.call.createPhoneCall({
  from_number: "+14155550100",
  to_number: "+14155550123",
  override_agent_id: agentId,
});
// -> { call_id, call_status }
Enter fullscreen mode Exit fullscreen mode

Watch the Enterprise wall: SSO, RBAC, a dedicated server and custom MSA, DPA or BAA terms all sit behind custom pricing. Pay-as-you-go gets a shared server and community support.

Bland AI

Bland AI pricing page showing Start at $0.14 per minute with no platform fee and 10 concurrent calls, Build at $0.12 per minute with a $299 monthly fee and 50 concurrent calls, and Scale at $0.11 per minute with a $499 fee and 100 concurrent calls

Bland refuses the itemized bill on principle: no token charges, no model pass-throughs. Start is $0.14 a minute with no platform fee, 10 concurrent calls and a 100 call daily cap, and signup includes two credits and an inbound number. Build cuts the rate to $0.12 for $299 a month with 50 concurrent calls and 2,000 daily. Scale is $0.11 for $499 with 100 concurrent and 5,000 daily. Minutes after a handoff to a human are billed separately at $0.05, $0.04 or $0.03.

Compliance is the strongest of the self-serve group: SOC 2 Type I and II, HIPAA eligibility, GDPR, PCI DSS, and a FedRAMP banner on the pricing page, though the signed BAA is an Enterprise item.

Do the arithmetic before you upgrade, though. Dividing that $299 fee by the two cents it saves you per minute means Build is a worse deal than Start until you are past roughly 15,000 minutes. And measured against Vapi's hosting fee, Bland costs about three times as much. Simplicity is the product here, and if your model bills are already small because you run cheap models on your own keys, it is a product you may not need.

ElevenLabs Agents

ElevenLabs is the voice model vendor selling the whole agent rather than just the speech. Its self-serve tiers bundle minutes instead of metering them:

Plan Monthly Minutes Concurrent
Free $0 15 4
Starter $6 75 6
Creator $22 275 10
Pro $99 1,238 20
Scale $299 - -
Business $990 - -

Going through the API instead, the Agents speech engine is $0.08 a minute and handles both STT and TTS in a single low-latency pipeline across more than 70 languages. Which model does the thinking is left to you.

If recognizable voice quality is the actual product, this is the shortest route to it, and the balance sheet is the deepest in the category at $781 million raised. Two things to keep in mind: the $0.08 excludes the LLM entirely, and bundled minutes evaporate faster than they read. Pro's 1,238 minutes is about 20 hours of talk time, which is not a call center.

LiveKit Agents and Pipecat

LiveKit Cloud pricing with Build at $0 per month, Ship from $50, Scale from $500 with HIPAA and security reports, alongside the GitHub badge showing the livekit/agents repository at 14.0K stars

Both open-source options let you own the pipeline, and both sell a cloud so you do not have to at first.

LiveKit Agents is Apache 2.0, 14,000 GitHub stars, Python and Node, with turn detection, telephony and native MCP tool support built in. Cloud pricing is itemized: Build is free, Ship starts at $50 a month, Scale starts at $500 and unlocks HIPAA plus security reports. The calculator's default phone stack adds up to $0.0672 a minute: agent session $0.0100, US number $0.0100, Gemma 4 31B $0.0014, Deepgram Nova-3 $0.0058, Cartesia Sonic 3 $0.0300, observability $0.0100. Switching the model to GPT-5 mini drops that LLM line to $0.0011.

Pipecat is BSD-2, 15,200 stars, maintained by Daily. Its abstraction is the frame: audio and text move down a chain of Python processors, one per stage, so transport, transcription, the model and speech synthesis are each a swappable link. Dozens of providers plug into those slots, OpenAI, Anthropic, Gemini, Deepgram, ElevenLabs and Cartesia included. Pipecat Cloud matches LiveKit's compute rate: the agent-1x profile (0.5 vCPU, 1 GB) is $0.01 a minute active, $0.0005 reserved, with no concurrency cap. WebRTC is free for one-to-one voice, PSTN dial-in and dial-out is $0.018, SIP runs $0.003 to $0.02, and Krisp noise cancellation is bundled up to 10,000 session minutes monthly. Model providers bill you directly.

Both carry the same tradeoff: what you get is a toolkit, not a finished product. Anything a managed platform hands you out of the box, cloned voices, staged rollouts, a dashboard to actually see what your calls are doing, becomes a line item on your own roadmap. If none of that appeals, Synthflow occupies the fully managed end of the market, where contracts open at $30,000 a year.

Price your own volume

Per-minute headline numbers hide six different bundles, so the only honest comparison is against your own traffic. This uses the rates above:

// voice-agent-cost.mjs - what a month of AI phone calls costs on each platform.
// Usage: node voice-agent-cost.mjs 20000   (minutes of talk time per month)
const minutes = Number(process.argv[2] ?? 20_000)

// Published rates, read from each vendor's pricing page on September 5, 2026.
const PLATFORMS = [
  { name: "Vapi (BYO model keys)", fixed: 0, perMin: 0.05, note: "models at cost, $0 if you bring your own API keys" },
  { name: "Retell (GPT-5 mini)", fixed: 0, perMin: 0.055 + 0.015 + 0.012, note: "infra + platform voice + LLM, all Retell-metered" },
  { name: "Bland Start", fixed: 0, perMin: 0.14, note: "all-in, 10 concurrent calls" },
  { name: "Bland Build", fixed: 299, perMin: 0.12, note: "all-in, 50 concurrent calls" },
  { name: "Bland Scale", fixed: 499, perMin: 0.11, note: "all-in, 100 concurrent calls" },
  { name: "ElevenLabs Agents", fixed: 0, perMin: 0.08, note: "speech engine only, LLM extra" },
  { name: "LiveKit Cloud (example stack)", fixed: 0, perMin: 0.0672, note: "calculator default: Gemma 4 31B, Deepgram Nova-3, Cartesia Sonic 3" },
  { name: "Pipecat Cloud (agent-1x + PSTN)", fixed: 0, perMin: 0.01 + 0.018, note: "STT/LLM/TTS billed by your providers" },
]

const usd = (n) => n.toLocaleString("en-US", { style: "currency", currency: "USD", maximumFractionDigits: 0 })
console.log(`\n${minutes.toLocaleString("en-US")} talk minutes per month\n`)
for (const p of PLATFORMS) {
  const total = p.fixed + p.perMin * minutes
  console.log(`  ${p.name.padEnd(32)} ${usd(total).padStart(9)}   ${p.note}`)
}
const breakEven = Math.ceil(299 / (0.14 - 0.12))
console.log(`\n  Bland Build beats Bland Start above ${breakEven.toLocaleString("en-US")} minutes/month\n`)
Enter fullscreen mode Exit fullscreen mode

At 20,000 minutes:

20,000 talk minutes per month

  Vapi (BYO model keys)               $1,000   models at cost, $0 if you bring your own API keys
  Retell (GPT-5 mini)                 $1,640   infra + platform voice + LLM, all Retell-metered
  Bland Start                         $2,800   all-in, 10 concurrent calls
  Bland Build                         $2,699   all-in, 50 concurrent calls
  Bland Scale                         $2,699   all-in, 100 concurrent calls
  ElevenLabs Agents                   $1,600   speech engine only, LLM extra
  LiveKit Cloud (example stack)       $1,344   calculator default: Gemma 4 31B, Deepgram Nova-3, Cartesia Sonic 3
  Pipecat Cloud (agent-1x + PSTN)       $560   STT/LLM/TTS billed by your providers

  Bland Build beats Bland Start above 14,950 minutes/month
Enter fullscreen mode Exit fullscreen mode

Do not read those totals without the notes. Vapi's $1,000 and Pipecat's $560 leave out the OpenAI and Deepgram invoices you still receive. Retell's and LiveKit's include them. Bland's include everything plus a compliance posture. Push it to 100,000 minutes and the gap runs from $2,800 to $14,000. The original article walks through the per-component tables each of these rates comes from.

If you are prototyping before you commit to a TTS bill, DevToolLab's Text to Speech tool will read your agent's opening line back in a browser voice, and the WAV File Generator produces silence or white noise for testing how your STT endpoint handles dead air.

Picking one

Shipping a product where voice is one feature and you already hold model keys, take Vapi for the itemized bill and the SDK. Want a single vendor metering everything with no keys to rotate, take Retell, and use the 20 free concurrent calls to load test. Regulated deployment where HIPAA, PCI or FedRAMP paperwork outweighs five cents a minute, take Bland, or Vapi Scale if you need model flexibility and can absorb $2,000 a month for HIPAA. Voice quality as the whole product at modest volume, ElevenLabs Agents. Engineers who want the cost curve under their own control, LiveKit Agents or Pipecat, on the vendor cloud until self-hosting pays off.

Three pricing shapes, then, and a funding round every few weeks to keep pushing them around. Every figure above is a September 5, 2026 snapshot with a short shelf life, so re-run the script yourself before you sign anything annual. The underlying question survives the repricing, though. Find whoever is metering your LLM minute, because that is the party holding your margin.

References

Top comments (0)