GatwAI

Vapi

Vapi connects telephony, speech models and business tools. Compare usage costs, retention, ZDR, engineering effort and practical voice-agent workflows.

Provider sources checked

What does Vapi do?

Vapi is a platform for developers integrating voice assistants into phone operations or their own applications. It connects speech recognition, language models, generated speech and telephony while allowing selection among supported providers. An agent can retrieve information or invoke approved functions; business logic and access permissions must fit the actual process. Its emphasis is programmable orchestration. This is useful when a team wants control over models, voices, integrations and evaluation. The trade-off is engineering and operational work: errors, interruptions, handoffs and external API failures need deliberate handling. One successful demonstration call is insufficient acceptance testing for a customer service.
Editorial research · 2026-11-14How we research →

Who is this tool for?

  • Engineering teams building phone or voice products.
  • Organizations controlling component choice and integrations themselves.

Strengths

  • Combine supported speech and model providers.
  • Connect custom functions and business systems.
  • Use simulations and log-based evaluation during development.

Limitations and requirements

  • Operation requires technical ownership.
  • Several cost components rather than one complete minute price.
  • Strict data minimization limits retrospective troubleshooting.

Who is behind the product?

Company
Vapi, Inc.
Headquarters
San Francisco, USA, according to its YC company profile.
Company founded
Y Combinator lists 2020 for the company rooted in Superpowered.
Product launch
The move to a voice API was publicly documented in 2023, separately from company founding.

A December 2024 company announcement describes Vapi as voice-agent infrastructure and confirms operation at that time.

Suggested ways to use it

Suggested workflows or verification steps. These are not hands-on tests; for unverified products they are not a recommendation to use the service.

Order status with restricted permissions

  1. Connect a test service containing fictional orders and read-only access.
  2. Restrict the agent to status queries; test identification, wrong numbers and outages.
  3. Compare call outcomes with the source and charges; test human handoff.

Goal: A proposed pilot with verifiable answers and no modification rights.

Compare models with identical calls

  1. Define identical tests for names, interruptions and slow integrations.
  2. Run two approved model combinations through the same scenarios.
  3. Compare completion rate, response delay and total cost rather than voice quality alone.

Goal: A traceable configuration decision before production.

Pricing, allowances and data

What should you check before choosing a plan?

Usage-based entry costs USD 0.05 per minute for Vapi hosting plus passed-through model, speech and applicable telephony charges. USD 5 starter credit is offered. Core adds USD 29 monthly; larger packages change support, concurrency and data controls. At 1,000 minutes, hosting alone is USD 50, not the entire bill.

What should you check about data and permissions?

By default, recordings, transcripts and other call artifacts can be stored. Entry pricing lists 14 days of raw-data retention. ZDR reduces stored conversation content while operational metrics remain, limiting some retrospective analysis. EU region and further controls depend on package. Check data flows and terms for every selected model and telephony provider.

External reviews and available evidence

Everett Butler describes a two-week personal test at Lindy, updated 9 October 2026. As a competitor's marketing head he has an explicit commercial interest; this is neither a neutral comparison nor a GatwAI test.

Reported positives

  • He highlights control over components.

Reported cautions

  • He notes technical setup effort and layered costs. Individual interface claims are not adopted without verification here.

These reports reflect the linked authors’ experiences. Test date, plan and task affect results; they do not establish a representative overall rating.

Product development and notable changes

  • Additional transcription providers

    The changelog adds Reson8 and Soniox stt-rt-v6 with conversational context. Its claimed quality impact was not measured here.

    Source for this change ↗

Which alternative fits which task?

Synthflow AI →

Compare an approach sold more explicitly as an enterprise deployment project.

Frequently asked questions

Is USD 0.05 per minute the final price?

No. Hosting is one component; selected speech, model and telephony services affect total consumption.

Does ZDR prevent all processing outside your organization?

No. It concerns retention, not the live processing route. Configuration of participating providers also matters.