A minute of AI on the phone costs five cents, seven cents, eight cents, eleven cents or fourteen cents, depending on which pricing page you open first. The pages are not wrong. They are measuring different things with the same unit.
Voice agent platforms sold themselves on per-minute rates through 2025 and 2026, and the market settled into four distinct billing models: a thin orchestration fee with model costs passed through, an itemized infrastructure-plus-menu model, an all-inclusive bundle, and annual enterprise contracts with no published rate at all. This piece takes each major platform's own pricing page, as published in November 2026, and puts them on comparable footing.
The four meters

Vapi: a platform fee with everything else at cost
Vapi's pricing page lists a $0.05 per minute hosting fee and nothing else included. Speech-to-text, the language model and text-to-speech are billed at provider cost with no markup, and $0 if you bring your own API keys. Telephony is itemized separately: Twilio inbound at $0.008 per minute, outbound at $0.014, Telnyx at $0.0055, while Vapi's own telephony, SIP and websockets are listed as free.
The page's own worked example says the honest range. For 1,000 minutes a month with Deepgram transcription, an OpenAI model and an ElevenLabs voice, Vapi estimates $82 to $129 total: $50 hosting, $10 transcriber, $8 to $45 model, $15 to $24 voice. The spread is almost entirely the model choice.
Success packages are priced separately: $29 a month for Core (10 concurrent calls, 5 numbers), a Pro tier at 10% of the hosting fee with a $999 monthly minimum (30 concurrent calls, 99% uptime SLA), and Premier with a 99.9% uptime SLA and a one-hour P0 response. Compliance is an add-on: $2,000 a month for HIPAA handling, $10 per month per extra concurrent line, $20 per extra organization.
Retell: infrastructure fee plus a published model menu
Retell's pricing page advertises $0.07 to $0.31 per minute, and the component table explains the width of that range. Retell Voice Infra is $0.055 per minute. Platform voices (Retell's own ElevenLabs Flash resales, Cartesia, OpenAI voices) are $0.015 per minute, while ElevenLabs v3 and v4 voices cost $0.08 to $0.10 per minute. The language model is billed per call minute from a long published menu: $0.0016 for GPT 5 nano at the bottom, $0.064 for Claude 5.5 Sonnet in the middle, $0.256 or more for GPT 6.2 Astra at the top. Speech-to-speech models like GPT Realtime 1.5 are a flat $0.345 per minute.
Two billing quirks on the same page matter at volume. Calls under 10 seconds that open with a dynamic greeting are billed as 10 seconds. And if your agent's average prompt (system prompt, tools, node prompts, transcript and retrieved knowledge) exceeds 4,000 tokens, billed time grows proportionally: a 4,800-token prompt makes a 60-second call bill as 72 seconds. Add-ons are metered too: knowledge base retrieval adds $0.005 per minute, PII redaction $0.01, safety guardrails $0.005, branded outbound caller ID $0.10 per call. The first 20 concurrent calls are free; beyond that, concurrency costs $8 per line per month. Your own SIP trunk or numbers are not charged; Retell numbers are $2 a month.
Bland: one number, models included
Bland's pricing page is the mirror image: "No token charges. No model-provider pass-throughs." Start is $0.14 per minute with no platform fee (10 concurrent calls, 100 calls per day), Build is $0.12 per minute plus a $299 monthly fee (50 concurrent, 2,000 calls per day), and Enterprise is contracted to volume with on-prem or VPC deployment available. The per-minute rate covers the language model, transcription and premium voices; telephony is billed separately at pass-through cost, on Bland's built-in Twilio or your own carrier, and bring-your-own-trunk customers skip transfer fees (warm and live transfers run $0.04 to $0.05 per minute otherwise). A 99.9% uptime SLA is listed on all plans. The compliance stack is stated on the page: SOC 2 Type I and II, HIPAA with a signed BAA, GDPR, PCI DSS self-attestation, ISO 27001, and FedRAMP 20x Class A certification.
The comparison is arithmetic. Bland's $0.14 all-in beats most itemized stacks only if your chosen model and voice on those stacks cost more than roughly nine cents combined. A Vapi build on small models can undercut it; the same build on a frontier model at $0.10+ per minute cannot.
ElevenLabs: minutes bundled into a subscription
ElevenLabs Agents pricing folds call minutes into tiers rather than charging pure usage: the $22 Creator tier includes 275 minutes, $99 Pro includes 1,238, $990 Business includes 12,375. Additional calls are $0.08 per minute on top. Bursting above your concurrency limit (up to 3x) costs double at $0.16 per minute. External LLM and telephony providers are billed separately, at cost. The model suits light-to-moderate call volume from a team that would otherwise pay for a standalone voice stack anyway.
Then there is the contract tier
Synthflow's pricing page states plainly that there is no free plan, no trial and no per-minute rate: it is sold only as an annual enterprise agreement scoped to call volume, telephony, integrations and launch support. Synthflow's own November 2026 comparison of ten enterprise platforms (a vendor document, read with that bias) records how the rest of that tier prices: Parloa per minute with no published rates, PolyAI custom for enterprise with self-serve from $49 a month, Sierra and Decagon priced per outcome or conversation with no public rates, Amazon Connect at $0.038 per voice minute with AI included plus telephony, Genesys Cloud from $75 per user a month plus AI usage. Synthflow says it handles over 1 million calls daily; its flagship case study has Batteries Plus running an agent inside Zoom Phone across 700-plus US locations.
What a 1,000-minute month actually costs

Third-party measurements put the abstractions to work. Stackbinary's 2026 per-minute analysis computes effective all-in rates near $0.13 to $0.14 per minute across Vapi, Retell and Bland at low volume, with the differences washing out at typical production stacks. CloudTalk's benchmarked run measured $0.0836 per invoiced minute on Vapi in July 2026 against $0.1408 on Bland: the thin-fee model won with a cheap model attached, and lost ground once add-ons like HIPAA ($2,000 a month) and numbers ($10 a line beyond ten) were priced in.
The pattern across both exercises: the advertised rate is a floor, a ceiling or a bundle, but rarely the invoice. What actually moves bills is the model choice (a 60x spread on Retell's menu), the prompt size (billed directly once past 4,000 tokens), compliance add-ons, and concurrency priced per line.
Picking a meter
- You want to own the model, voice and transcriber choices, or bring your own keys: Vapi or Retell. Vapi if you want raw pass-through and control-plane tooling; Retell if you want the model menu priced per minute up front and flow-building features as add-ons.
- You want one predictable number and a closed, compliance-forward stack: Bland.
- You are a contact center buying a deployment, not an API: Synthflow, Parloa or PolyAI, on contracts you will negotiate anyway.
- You want ElevenLabs voices and modest volume: ElevenLabs Agents bundled minutes.
- You are already on AWS and want AI minutes inside the suite rate: Amazon Connect.
The five-cent minute and the fourteen-cent minute are increasingly the same product at the same volume. The bill is decided by everything the headline rate excludes.
