Vapi

What it is

Vapi is developer infrastructure for AI phone agents. It orchestrates the real-time loop — speech-to-text, the LLM, text-to-speech, turn-taking, interruptions, and the phone line — and lets you choose every component in that loop yourself. You configure an assistant through the dashboard or API, attach tools, point it at a number or SIP trunk, and pay Vapi $0.05 per minute for hosting while the model, voice, and transcription providers bill at their own cost. It sits in the build-it-yourself layer next to Retell AI, not the managed-service layer of PolyAI or Parloa.

The San Francisco company started as Superpowered, a YC-backed meeting-notes app, and pivoted to a voice API in late 2023 under CEO Jordan Dearsley. On 12 May 2026 it announced a $50M Series B led by Peak XV, with M12, Kleiner Perkins, and Bessemer participating, at a reported valuation around $500M and $72M raised in total. Vapi says it has passed 1 billion calls and 1 million developers, and that enterprise ARR grew tenfold since its previous round. The headline customer is Amazon Ring, which evaluated more than 40 voice vendors and now routes all of its inbound support calls through Vapi; Ring’s VP of software development said the team went from zero to production in two weeks.

Why it shows up in CX and support stacks

• Every component is swappable. Deepgram or another transcriber, OpenAI, Anthropic, or Google models, ElevenLabs or Vapi’s own voices, Twilio, Telnyx, Vonage, or your SIP trunk. When a cheaper model or a better voice ships, you change one setting instead of migrating platforms. Models are billed at provider cost with no markup.
• Multi-agent calls through Squads. A handoff tool passes a live call from one assistant to another with context and extracted variables, so an authentication agent, a billing agent, and an escalation agent can each stay small and testable.
• Composer builds the first draft. An in-dashboard assistant, launched in alpha in early 2026 and improved in August, turns a plain-language description into an assistant with numbers, tools, and integrations. Vapi reports Composer users are about 3x more likely to reach 100 call minutes — a vendor figure, but the direction matches what onboarding data usually shows.
• MCP in both directions. An official MCP server at https://mcp.vapi.ai/mcp lets Claude or another client list and create assistants, place or schedule calls, and pull call records. Separately, an MCP tool lets a running voice agent call your own MCP servers mid-call.
• Testing is in the box. Simulation suites (August 2026) run scripted callers against an assistant before deploy, on every plan.

Pricing reality

The $0.05 hosting fee is the smallest line on the bill. Vapi’s own calculator prices 1,000 minutes at $82–$129 a month: $50 hosting, about $10 transcription, $8–$45 for the LLM, and $15–$24 for an ElevenLabs voice. Add telephony — Twilio charges $0.008/min inbound and $0.014/min outbound, Telnyx $0.0055/min, Vapi’s own SIP is free — and a mainstream build lands at roughly $0.09–$0.14 per minute all-in. A frontier model or premium voice pushes past that.

Then the plan layer. Pay-as-you-go caps you at 4 concurrent calls, 1 number, 14-day data retention, and Discord or email support with no SLA. Core is $29/month for 10 concurrent calls and a 2-business-day support SLA. Pro is 10% of hosting with a $999/month minimum for 30 concurrent calls, Slack support, and a 99% uptime SLA; Premier is custom with a 99.9% SLA and 1-hour P0 response. Extra concurrency is $10 per line per month. HIPAA is a $2,000/month add-on that includes the BAA.

At 100,000 minutes a month on a mainstream build, expect about $9K–$14K in usage plus the $999 Pro floor — roughly $120K–$180K a year, before HIPAA. That is in the same band as Retell at volume; Vapi is cheaper for small builds and less predictable for large ones.

Best for

CX-ops and support-engineering teams with engineers who will own a voice agent like any other production service — building inbound support, triage, scheduling, or outbound reminder lines from about 10,000 to several million minutes a month — and who want to pick the model, voice, and carrier themselves rather than accept a bundle.

Skip it if nobody on the team can write a webhook, if you need a vendor to design and run the agent, or if you are a small clinic that needs HIPAA on a budget: $2,000/month buys a lot of minutes elsewhere.

Versus the alternatives

Retell AI is the other default in this layer. It bundles voice infrastructure, platform voices, and telephony into one meter (about $0.15/min on a mainstream build), lets you sign the BAA online, and scores calls with built-in QA. Pick Retell when you want one predictable rate and HIPAA without a $2,000 monthly add-on; pick Vapi when your engineers want to choose and swap every component, or need the larger integration surface and MCP server.

ElevenLabs Agents is the fastest-growing entrant, built on ElevenLabs’ own voices — pick it when voice quality and language coverage decide the project. Synthflow is the no-code option, for teams where the builder is not an engineer. If you want a vendor to own delivery, buy managed: PolyAI, Parloa, or Sierra.

If none fit, spend Vapi’s $5 free credit on your single highest-volume intent and compare its measured cost per resolved call with your current human cost per call before signing anything.

Watch-outs

• The bill has four or five meters. Hosting, transcription, model, voice, and telephony each move independently. Guard: build the exact stack in the pricing calculator, run a pilot, and forecast from measured average handle time and per-call cost in the call logs — not from the $0.05 headline.
• Fixed fees change under you. A customer on Vapi’s support forum reported the HIPAA add-on doubling to $2,000/month without notice and a disputed $1,000 charge that went unanswered for eight days. Guard: get HIPAA, concurrency, and success-package pricing written into an order form with a notice period before production traffic depends on it.
• Pay-as-you-go support is thin. Community posts describe 2-3 day gaps between replies, and the default plan has no SLA. Guard: move to Pro or Premier before the line carries revenue-critical calls, and keep a human fallback number in your IVR for outages like the 1-hour-55-minute SIP incident on 14 April 2026.
• Default concurrency is low. Four simultaneous calls on pay-as-you-go will drop traffic during the first spike. Guard: load-test your peak hour in simulation and buy concurrency at $10/line before launch, not after the first busy signal.