Vapi
4.2
Vapi is infrastructure for voice agents that answer and place real phone calls. It handles the parts that make voice hard — speech recognition, turn-taking, interruption, latency, telephony — and leaves you the prompt and the business logic. Billing is per minute, so a pilot costs what the calls cost rather than an annual contract.
- Commercial use OK
- Free tier
- Enterprise contract
- Pricing
- 従量課金
- Free tier
- Free tier
- Commercial use
- Permitted
- Japanese UI
- Not confirmed
- Enterprise contract
- Available
What it is good for
- Handling routine inbound calls that currently queue for a human
- Appointment booking, confirmation and reminder calls
- Qualifying inbound leads before a person picks up
- Testing whether voice automation works at all, cheaply
Strengths and trade-offs
Strengths
- Per-minute pricing: a pilot costs tens of dollars, not a contract
- Handles interruption and turn-taking, which is where naive builds fail
- Swappable speech and language models rather than one fixed stack
- Telephony included, so there is no separate carrier integration
Trade-offs
- Latency is the whole product; a poor configuration sounds broken immediately
- Call recording and consent rules differ by jurisdiction and are yours to handle
- Voice quality in Japanese lags English and needs its own evaluation
Editors' review
The practical way to put an AI agent on the phone without building telephony
Building a voice agent from parts means solving recognition, turn-taking, barge-in, latency and telephony before you get to the actual conversation. Vapi supplies that layer, which moves the project from months to days and makes the question 'is this worth doing' answerable.
Per-minute pricing matters for the same reason. A pilot on a single phone number costs what the calls cost, so a team can run real traffic through it and listen to recordings before anyone signs anything.
The failure mode is latency. A pause of a second or two reads as a broken line to the caller, and the model, voice and routing all contribute. Japanese voice quality also needs separate evaluation — it is behind English — and recording consent is a legal question that the platform does not answer for you.
Verdict
The right foundation for a phone-answering agent. Budget the pilot for tuning latency, and settle the consent question first.
Compiled from official documentation and public sources, with our own editorial judgement. We do not publish figures we cannot source.