If you have priced a voice agent in 2026, you have run into the same problem: every platform quotes a per-minute rate, and no two quotes measure the same thing. Vapi says five cents. Retell says it starts at seven. Bland says somewhere around eleven, flat. A buyer comparing those three numbers on a spreadsheet would conclude Vapi is the cheapest by a country mile — and would be wrong, because two of those three numbers leave out the speech recognition, the language model, and the voice itself.
The stakes are higher this year than last. The phone is now the highest-intent channel a service business has, and the data on what an unanswered call costs is no longer soft. Independent SMBs miss or abandon 12% to 50% of inbound calls, rising to 48–55% during demand surges, per Ascero’s 2026 SMB Phone-Leakage Report. In real estate, 73% of buyers hire the first agent who responds (NAR, 2024). In the trades, a single missed install call can be a $5,000–$25,000 opportunity. Meanwhile the infrastructure layer has commoditized: xAI shipped a no-code Voice Agent Builder at roughly six cents a minute all-in, and ElevenLabs cut conversational-AI pricing to 8–10 cents a minute. Building a voice agent is no longer the hard part.
Which means the buying decision has moved. It is no longer can we automate the phones? It is what does the stack actually cost to own, and who carries the compliance risk? Here is the honest 2026 comparison of the three platforms that come up in every evaluation, and the shortcut agencies are quietly taking instead.
Vapi’s bet is that voice agents belong in code. You choose your own speech-to-text, LLM, and text-to-speech providers, wire in your tools, and Vapi handles real-time orchestration, telephony, and call state. That flexibility is exactly why every other vendor benchmarks itself against it — and exactly why its pricing page needs a footnote.
The published platform fee on Vapi’s pricing page is $0.05 per minute. Provider costs for STT, LLM, and TTS are passed through at cost on top of that. In practice, most teams land between $0.07 and $0.25 per minute, and pass $0.30 with premium voices and large context windows. Then comes the Success Package: $29 per month for Core, or a $999 per month minimum for Pro if you want a one-business-day SLA and 99% uptime. HIPAA handling is a separate $2,000 per month add-on. Zero data retention is another $1,000 per month.
None of that is unreasonable. All of it is invisible in a two-line quote, and every model swap you make changes the bill.
Retell bundles the choice into one line item. Rates on Retell’s published pricing run from about $0.07 to $0.31 per minute depending on voice and model, with voice infrastructure included, plus $2 per month per phone number and 20 concurrent calls. A mainstream configuration lands near $0.11–$0.13 per minute — a number you can actually forecast. It also puts HIPAA, SOC 2 Type 1 and Type 2, and GDPR on self-serve at no extra fee, which Vapi gates behind enterprise tiers or bills separately.
The catch is not price, it is posture. Retell is a low-code developer platform. Someone on your team owns the agent configuration, the prompt iteration, the knowledge base, and the testing loop. Retell even shipped Conductor, a production review interface, which is a quiet admission that agents fail in ways demos do not reveal.
Bland sells simplicity on a single flat rate: roughly $0.11–$0.14 per minute covering LLM, STT, and TTS with no separate token charges, plus a platform fee, with bundled numbers and HIPAA eligibility under a BAA. One number, one invoice.
The number to interrogate is latency, because latency is what makes a caller hang up. Bland advertises approximately 400ms, with methodology undisclosed. Independent testing measured 850ms median and 1,180ms p95 — the largest gap between claim and measured result in the category. Retell’s claim and independent test roughly agreed; Vapi hit 720ms median against a sub-500ms target.
Every one of these platforms is really competing against one number: what a human costs to do the same job. A fully loaded US onshore call-center agent runs about $0.43 per minute; nearshore runs $0.20–$0.30; offshore runs $0.10–$0.23. A traditional live answering service at the entry tier is $5.00 per minute. Against those baselines, ten cents a minute looks like a rounding error — and that is precisely why the per-minute debate is the wrong debate.
| Cost center | What the vendor quote shows | What you actually pay |
|---|---|---|
| Speech stack (STT + LLM + TTS) | Often excluded | $0.03–$0.25/min and up |
| Telephony and numbers | Sometimes bundled | $1–$2/number + carrier minutes |
| Engineering and prompt maintenance | Never shown | Your team’s hours, every month |
| Compliance (HIPAA, retention, consent) | Enterprise add-on | Up to $2,000–$3,000/month |
| CRM, portal, and booking glue | A separate tool | Another subscription plus integration work |
Read that table again and the headline rates stop mattering. The real question is how many vendors you are reconciling, and how much of the build you own after launch.
While the platforms were racing on price, the regulatory floor moved. The FCC has confirmed that an AI-generated or cloned voice counts as an “artificial voice” under the TCPA, which means an outbound AI voice call needs the same prior express consent as a prerecorded robocall, plus clear identification and a working opt-out. Penalties run $500 to $1,500 per call with no aggregate cap.
State law is stacking on top of it. California has required AI-voice disclosure on calls since January 1, 2025, with reported penalties of $500 per violation. Texas’ Responsible AI Governance Act took effect January 1, 2026. Washington signed a companion-chatbot disclosure law in March 2026, Colorado’s AI Act is live, and Utah requires disclosure in high-risk consumer interactions.
The disclosure obligation falls on the business operating the agent — the contractor, the clinic, the agency — not on the toolkit it was built with. A two-minute setup does not ask which states your callers live in.
This is the part that a Vapi-versus-Retell spreadsheet cannot answer, and it is where a lot of DIY deployments quietly become liabilities. Consent flows, disclosure scripts, opt-out paths, and recording-handling policy are not features you bolt on at the end. They are the reason a self-serve platform with an inherited compliance posture often wins over a cheaper per-minute rate.
Build on a developer-first platform when at least three of these are true:
If two or more of those are false, you are not buying a voice agent. You are buying a permanent engineering obligation with a per-minute meter attached.
Most agencies selling AI receptionists to local businesses never touch Vapi or Retell directly. They buy one all-in-one platform, reskin it with their own brand and domain, and sell it as their product. The margin math is not subtle: wholesale platform cost is a flat monthly fee, the client pays $197–$597 a month, and the agency keeps the spread. The technology risk, the speech stack, the telephony, and the compliance backbone stay with the platform.
That is the model behind Nida. It is not a voice-agent vendor that happens to have a dashboard — it is a white-label business platform where the AI voice agent, the CRM, the client portal, booking, and the website live in one workspace. For an agency that already resells websites, adding voice is a line item, not a new stack: the same branded portal the client already logs into now shows call recordings, transcripts, and booked jobs. We walk through the full reseller economics in how to resell AI voice agents for recurring revenue, and the running-cost breakdown of the platform-by-platform path in what an AI voice agent actually costs to run.
The reason agencies consolidate rather than stack is the same reason the per-minute comparison misleads: every additional tool adds a login, an integration, a support surface, and a place for the lead to fall through. A lead captured by the phone agent but living in a different system from the CRM is a lead somebody has to move by hand. Bundling websites and portals into one retainer is the version of this that survives contact with a client who is actually paying attention.
Consider an agency running 20 local clients, each taking roughly 600 calls a month at four minutes — 48,000 minutes total.
At $297 a month across 20 clients, that is $5,940 in monthly recurring revenue from a product you did not build, against a platform cost that starts in the low hundreds. The first one or two clients cover the platform entirely. The rest is margin.
The catch to be honest about: white-label is not free money. You own the client relationship, which means you own the onboarding, the escalation, and the renewal conversation. What you do not own is the speech stack, the telephony, the model-risk, or the compliance scaffolding — and in 2026, that is the expensive half.
None of this makes Vapi, Retell, or Bland bad products. Vapi is the reference implementation for teams that want to own the stack. Retell is the fastest credible path for a technical team in a hurry. Bland is the simplest invoice in the category, if the latency holds on your calls. They are the right answer to a narrow question: how do we build and run a voice agent ourselves?
For an agency, a service business, or anyone whose actual product is the customer relationship, the better question is a different one. If the pitch is “AI receptionist, booked jobs, one portal, your brand, flat monthly fee,” then the per-minute rate was never the thing you were buying in the first place.
None is reliably cheapest, because the headline rates measure different things. Vapi’s $0.05 per minute is hosting alone and excludes speech-to-text, the LLM, and text-to-speech, putting most real setups at $0.07–$0.25 and up. Retell publishes $0.07–$0.31 with its voice infrastructure included, landing near $0.11–$0.13 for a mainstream configuration. Bland is a flat $0.11–$0.14 plus a platform fee. Model your own call volume, not the headline number.
Retell is low-code and Vapi is developer-first, but both expect someone to own the agent configuration, prompt iteration, knowledge base, and testing loop after launch. Vapi defines agents in JSON with bring-your-own-key provider billing. If nobody on your team will own that permanently, a managed or white-label platform is the cheaper long-term decision even at a higher sticker price.
At 600 calls a month per client and four minutes per call, 20 clients is roughly 48,000 minutes monthly. Directly on a developer platform at $0.13–$0.20 all-in, that is $6,240–$9,600 in usage alone, before add-ons and engineering time. On a flat white-label platform with sub-accounts, the platform fee starts in the low hundreds per month and you set your own per-client pricing.
Inbound answering is straightforward. Outbound is regulated: the FCC treats AI-generated voices as “artificial voices” under the TCPA, so outbound AI calls need prior express consent, clear identification, and a working opt-out, with penalties of $500–$1,500 per call. California requires AI-voice disclosure on calls, Texas’ AI governance act took effect January 1, 2026, and Washington, Colorado, and Utah have disclosure rules. The obligation sits with the business operating the agent, not the tool vendor.
Yes. White-label platforms let agencies run the AI receptionist, portal, emails, and billing under their own domain and logo, usually on a flat monthly fee with a sub-account per client. Partners commonly charge $197–$597 per client per month, and one or two clients typically cover the entire platform cost.
If you are buying voice for your own business, insist on one number: all-in cost per answered call across platform, AI, telephony, and upkeep. If you are an agency, the better move is to skip the component stack and sell a branded outcome — answered calls, booked jobs, and one portal your client logs into. See how Nida prices a white-label workspace with the voice agent, CRM, booking, and client portal built in.
← Back to the blog