What is Vapi?

Vapi is a developer-first platform for building, testing and deploying real-time AI voice agents that talk to people over the phone or in a browser. It sits between your phone system and your AI models, handling the hard parts of a live conversation — streaming audio, detecting when the caller has finished speaking, managing interruptions, and passing text back and forth between speech recognition, a language model and a voice synthesiser fast enough that the whole thing feels like a phone call rather than a form.

The company was founded in 2020 by Jordan Dearsley and is based in San Francisco. It has raised in the region of $75m across two rounds from investors including Bessemer Venture Partners and Peak XV Partners, which puts it among the better-funded independent players in a crowded field. Its closest comparisons are Retell AI, Bland and Synthflow, and — at the model layer rather than the orchestration layer — ElevenLabs, whose voices many Vapi deployments actually use.

The important framing: Vapi is not a finished product. It’s a development framework. That single fact explains almost every strength and every complaint below.

Key features

Total model flexibility — the core selling point

Vapi’s defining feature is that you’re not locked into anyone’s stack. Language model from OpenAI, Anthropic, Groq or an open-source option; speech-to-text from Deepgram, Google or Whisper; text-to-speech from ElevenLabs, PlayHT, Deepgram or Azure. Every component is swappable independently, which matters more than it sounds — when a cheaper or faster model ships, you change a setting rather than migrate a platform. No competitor offers this breadth of choice at every layer.

API-first design and SDKs

Everything in Vapi is reachable programmatically. You can create agents on the fly, configure call logic dynamically, and wire the whole thing into your existing backend without touching a visual builder. There are SDKs for embedding voice directly into web and mobile apps, plus CLI tooling for scripted test calls. For teams building voice into a product rather than bolting it onto a business, this is the whole point.

Flow Studio, Blocks and Squads

Vapi has softened its developer-only edges. Flow Studio is a drag-and-drop visual builder for mapping multi-step conversations, built from components called Blocks that control what the agent does at each stage. Squads takes it further, letting several specialised agents hand off to one another — one qualifies the lead, another books the appointment. It works, though reviewers consistently rate the builder as less polished than the no-code specialists.

Telephony, widgets and integrations

You can buy numbers directly through Vapi (US only) or bring your own from Twilio, Telnyx or Vonage, which is how most non-US deployments run. There’s now an embeddable web widget with decent customisation, so you can put a voice agent on a site without developer time, plus a text-chat mode added after user demand. Integrations number around 40 — HubSpot, Notion, Clay, Zapier and the usual suspects — which is respectable but well short of a packaged platform’s library.

The standout: swappable infrastructure

The thing Vapi does that nothing else quite matches is refusing to make the expensive decisions for you. Every other platform in this category bundles a stack and marks it up. Vapi passes provider costs through at cost and charges for the orchestration layer alone. That’s why the base rate is the lowest available, and why sophisticated teams can genuinely run agents cheaply — if you pair budget speech-to-text with an efficient model and a mid-tier voice, you can sit near the bottom of the cost range while a competitor’s customer pays a flat rate for a stack they didn’t choose. The flexibility is the product.

Scorecard

Overall score: 7.3 / 10 — the average of the eight categories above.

Pricing

There are no seat fees and no subscription. Vapi charges a platform fee per minute and passes every other cost through at provider rates. This is honest in structure and genuinely hard to forecast in practice.

The honest read: the cheapest headline rate in the category and rarely the cheapest final bill. New accounts get $10 in free credits, which is enough to build and test something real. Pay-as-you-go caps you at 10 concurrent calls — a limit that arrives faster than people expect on a busy inbound line, and half what Retell AI includes — and lifting it means an enterprise conversation, where published estimates for annual budgets sit in the $40,000–$70,000 range. Multiple sources converge on $0.30-plus per minute for premium configurations, roughly six times the advertised rate. None of this is hidden exactly; it’s just that the number in the marketing and the number on your invoice are different things, and you should model the second before committing.

Pros & cons

Honest weaknesses

The recurring complaint is cost surprise. The $0.05 figure is accurate for what it covers and misleading about what you’ll pay, and because billing spans multiple vendors you’re reconciling several invoices to work out what a call actually cost. Teams that budget from the headline rate consistently underbudget, and there’s no cap unless you set one — a volume spike becomes a bill spike with nothing in between.

Latency deserves scrutiny too. Sub-500ms is quoted everywhere and is achievable with an optimised configuration, but independent testing more commonly lands in the 550–800ms range, with some reporting 800–1200ms on default pipelines and occasional outliers of several seconds. On a voice call that difference is audible. If response speed is critical, test with your own stack in your own region before you commit — and note that extremely low latency brings its own problem, since an agent that responds instantly tends to talk over people.

Then there’s the support model. Below enterprise, help means the documentation and a Discord community, with ticket responses reported as taking days. The docs are good, but self-service troubleshooting suits experienced developers and punishes everyone else. There are also scattered reports of agents breaking after platform updates — manageable if you have someone monitoring, painful if the agent is answering your phone unattended. Vapi’s testing tooling is also thinner than Retell’s, which makes catching those regressions before customers do harder than it should be.

One editorial note worth stating plainly: a large share of the Vapi reviews you’ll find online are published by competing voice platforms, and they lean predictably. We’ve weighted independent testing and user forums more heavily than vendor blogs here, and we’d suggest you do the same when researching this category.

Who is Vapi for?

Vapi is the right pick for engineering teams embedding voice into their own product, agencies building custom agents for clients, and technical operators who want to optimise the stack themselves. If component swappability matters, if you have workflows too complex for a visual builder, or if you’re running high call volumes where a few cents a minute compounds into real money, nothing else gives you this much room to work.

It’s the wrong pick if you’re a small business owner without developer access. The setup is not a weekend project, the support won’t carry you, and the cost model rewards people who understand every layer of it. For a takeaway wanting to stop missing calls or a trades business chasing no-shows, a packaged platform — or the practical approach in our appointment reminder guide — will get you further, faster, for less. Vapi rewards expertise. Bring some, or bring someone who has it.

FAQ

AI voice agentVapivoice AIphone automationAI tool review