Vapi

What Is Vapi?

Vapi is a developer platform for voice agents that answer and place real phone calls. It does not hand you a finished agent. It hands you the plumbing that turns three separate models — speech recognition, a language model and a speech synthesizer — into something capable of holding a conversation without the pauses that make a caller hang up.

That plumbing is most of the work, and it is the part people underestimate. A voice agent has to decide when the caller has stopped talking, stop itself the moment they start again, keep the gap between question and answer short enough that the silence does not read as a dropped line, and know when to give up and find a human. None of that is a prompt engineering problem. Vapi owns the turn-taking, the audio streaming, the telephony and the call state, and leaves every model choice to you.

The company crossed a billion handled calls in May 2026, alongside a $50 million Series B led by Peak XV at a reported valuation near $500 million, with Microsoft's M12, Kleiner Perkins and Bessemer taking part. The reference customer it is asked about most is Amazon Ring, which evaluated more than forty vendors before routing its inbound support traffic through the platform.

Как это работает

You Choose Every Model in the Stack

Vapi is deliberately unopinionated about models. Transcription can come from Deepgram or Gladia, reasoning from OpenAI, Anthropic or Google, the voice from ElevenLabs or a dozen other providers. Any layer can be swapped without rewriting the agent, which matters in a field where the best-sounding voice model changes every few months.

The platform targets sub-600 millisecond response times, and hitting that number is a function of your choices as much as theirs. A slow reasoning model or a chatty system prompt will blow the latency budget, and Vapi will let you do it. Control cuts both ways: nothing stops you from assembling a stack that sounds worse than a managed competitor's default.

Assistants, Squads and Workflows

The simplest unit is an assistant: one system prompt plus a set of tools. It covers the obvious cases — qualifying a lead, booking an appointment, answering the questions a business answers forty times a day.

A squad is several assistants that pass the same call between them while keeping the context, which is how you build triage without one enormous prompt trying to be a receptionist, a scheduler and a support tier all at once. Workflows sit at the other end for calls that must follow a fixed path — an identity check, a payment confirmation — where a language model improvising is a liability rather than a feature.

The distinction is worth understanding before you start, because prompt-only agents degrade in a predictable way: every new rule you bolt on makes the earlier rules a little less reliable. Splitting the job across a squad is the fix, and retrofitting it later is more work than starting there.

Tools Are Where the Value Sits

Tools let the agent call your API, query a database or write to a calendar in the middle of a conversation. This is the line between a demo and a product. An agent that only talks is a novelty; one that can look up an order, check availability and change a booking is doing a job somebody used to be paid for.

Calls reach the agent by phone number, through a web SDK embedded in your own product, or over SIP if you already have carrier arrangements you would rather keep. There is also a CLI, which tells you something about the intended audience.

Сколько это стоит

Vapi charges for orchestration and passes model costs through at cost. That makes the platform fee easy to reason about and the total bill harder, because the total depends on which models you picked.

In practice a straightforward English-language support agent lands somewhere between eight and fifteen cents a minute all in. The spread comes almost entirely from the language model, so the single biggest cost lever you control is not the vendor — it is whether the job actually needs a frontier model.

Кому это даёт больше всего

Engineering Teams That Want the Internals

If your instinct on hearing that a platform picked the speech model for you is to ask which one and why, this is the platform for you. Everything is addressable, everything is swappable, and the documentation assumes you can read an API reference.

Companies With Telephony They Cannot Move

Bring-your-own-carrier over SIP means the numbers, contracts and routing you already have stay where they are. For anyone with an established contact centre this removes the objection that usually kills a voice AI pilot before it starts.

High-Volume Inbound Support

The Amazon Ring deployment is the proof point the company leans on, and it is the right shape of workload: predictable questions, enormous volume, seasonal spikes that would otherwise require hiring. Per-minute pricing with no platform fee suits traffic that is lumpy rather than steady.

Products That Embed Voice

Because the web SDK drops into your own application, Vapi is a common choice for software companies adding a voice layer to something they already sell, rather than for businesses buying an agent to answer their own phone.

На что обратить внимание

• You assemble the bill yourself. The $0.05 platform rate is the smallest line on the invoice, and model costs can quietly triple it.
• It is a developer platform, not an application. There is a dashboard, but a non-technical team will not get a production agent out of it alone.
• Latency is your responsibility. The sub-600ms figure describes what the orchestration allows, not what your particular stack will deliver.
• Compliance costs extra and costs a lot. HIPAA at $2,000 a month is a real decision for a small team, and it is not negotiable downward on the self-serve plan.
• The free allowance is a test drive. Sixty minutes will validate that the concept works and tell you nothing about how the agent behaves at scale.
• Voice agents fail in public. A text chatbot that misfires is a bad message; a phone agent that misfires is a customer telling the story at dinner.

Частые вопросы

Is Vapi free?

No. There is a free allowance of just over an hour of calls for evaluation, after which you pay $0.05 a minute for the platform plus whatever the models and telephony cost. There is no free tier that survives contact with real traffic.

Do I need to write code?

For anything beyond a prototype, yes. The dashboard will get an assistant answering the phone, but tools, integrations and the error handling that makes an agent trustworthy are code. Teams without an engineer are better served by a no-code platform.

How is it different from Retell AI or Bland?

Vapi orchestrates models you choose; Bland runs its own stack end to end and sells the reliability that comes from owning it; Retell sits between the two with a visual flow builder over a similar mix-and-match architecture. Choose Vapi for control, Bland for regulated workloads where the vendor owning the whole path matters, Retell if you want the middle.

Can it use my existing phone numbers?

Yes. You can buy numbers through the platform or connect your own carrier over SIP, which is usually the deciding factor for companies with an existing contact centre.

Is it suitable for healthcare or finance?

It can be. HIPAA support and zero data retention are available as paid add-ons at $2,000 and $1,000 a month respectively, and the Scale plan covers SOC 2, PCI, SSO and data residency. Budget for compliance as a line item rather than assuming it is included.

What happens when the agent cannot help?

You define the handoff. Calls can transfer to a human, to another assistant in a squad, or to a fixed workflow. Designing that escape hatch carefully is the difference between an agent people tolerate and one they resent.

Итог

Vapi is the voice platform for people who want to see the wiring. It gives you the hard parts of real-time phone conversation — latency, interruption, telephony, state — and then gets out of the way on every decision that shapes how the agent sounds and what it costs.

That freedom is the whole pitch and the whole caveat. A team with an engineer who enjoys this will build something better and cheaper than any managed alternative. A team without one will build something slow, spend more than they expected, and conclude that voice AI does not work. Know which team you are before you start.