Vapi Review 2026: The Cheapest Way to Sell AI Phone Agents

I have run Vapi agents for two clients since early 2026 - a dental office that wanted after-hours booking and a property manager who needed outbound lead qualification. The platform fee is real and it is low. The invoice that lands at the end of the month is not the number on the pricing page, and that gap is the first thing anyone building a voice-agent business has to understand.

What Vapi Actually Is

Vapi is not a phone system and not a chatbot. It is the orchestration layer that sits between four vendors you pick yourself: a speech-to-text engine (Deepgram, AssemblyAI), an LLM (GPT, Claude, Gemini), a text-to-speech voice (ElevenLabs, Cartesia) and a telephony provider (Twilio, Vonage). You describe the agent - system prompt, tools, voice, transfer rules - and Vapi handles turn-taking, interruption handling, latency and routing. A well-tuned agent answers in under a second. There is an API, SDKs, and a dashboard for logs and transcripts.

That architecture is the point. Nothing is locked in. You can swap a $0.30/min voice for a $0.05/min one, or move from a frontier model to a small one, and watch the per-minute cost move the same day. It is also the source of the pain: four moving parts means four invoices and four places for latency to creep in.

The Money Angle

Small businesses hate missed calls, and the numbers are brutal. A missed call at a dental office is a $300-800 patient, at a law firm a $1,000+ consult, at an HVAC company a $500 job. That pain is why people are selling AI receptionists as a service.

What It Really Costs

The published fee is $0.05 per minute, billed to the second, and that is only Vapi's orchestration cut. Everything else is passed through at cost: STT about $0.005-0.008/min, TTS about $0.05-0.36/min depending on voice, telephony about $0.0085-0.014/min plus about $2/month per number, and the LLM billed per token - which Vapi's own docs call the largest cost you control, driven by how long your system prompt is and roughly five model calls per conversation turn.

Real all-in spend lands around $0.10-0.30 per minute, and roughly $0.12-0.24/min on a mid-tier voice before tokens. That means 500 minutes is about $75-125 a month, 2,000 minutes about $300-500, and 10,000 minutes about $1,500-2,500. Bring your own API keys and Vapi charges $0 for those line items - you just pay each vendor directly, which is cheaper but more admin.

The extras are where budgets break. HIPAA compliance is a $2,000/month add-on and Zero Data Retention is $1,000/month, on top of usage. Extra concurrent lines beyond the included 10 cost $10 each per month. SMS follow-ups are $0.005 per message. Build keeps call recordings for only 14 days, which is a problem the first time you want to review a month of calls. Scale is an annual contract with quote-only pricing, so you cannot budget it in a spreadsheet.

Where It Falls Apart

Latency. Users consistently flag it as unpredictable, and on a voice call that is not a cosmetic issue - a 1.5-second pause makes a bot feel broken. You fix it by tuning the stack, not by clicking a setting.

The dashboard is dense. Reviewers on G2 (a 4.2, off only a handful of reviews) describe it as overwhelming and ask for a simple mode. This is a tool for someone who reads logs and understands token costs, not for a solo operator who wants to hand a client a login.

And there is no compliance floor. Build ships without SSO, RBAC or SOC 2, so most mid-size companies will fail it in a security review. That pushes you to Scale, where the price is whatever the sales call says it is.

Who Should Pay

Buy Vapi if you have engineering time, want to sell voice agents as a service, and care about the per-minute math - it is the cheapest orchestration layer you can build a business on, and the free trial credit lets you prove it before you commit.

Skip it if you want a turnkey no-code product, if you need HIPAA or SOC 2 from day one without a $2,000/month add-on, or if nobody on your team enjoys watching latency dashboards. For those cases, a bundled platform like Bland at a flat $0.14/min or Synthflow at about $0.08-0.09/min will cost more per minute and save you far more in headaches.