Vapi

What Vapi actually meters

Vapi meters a connected call minute and splits it into five lines: hosting, transcriber, model, voice and transport. Only hosting is Vapi's revenue, and the page says the rest bills "at the provider's listed price with no markup".

Transcription bills two audio channels. Vapi transcribes caller and assistant both, so one conversation minute buys two minutes of speech to text, and that doubling explains the rates. Our Deepgram entry puts Nova-3 streaming at $0.0048 a minute; two channels is $0.0096, squarely inside Vapi's $0.00953 to $0.00994 Deepgram range. AssemblyAI streams at $0.0025; two channels is $0.0050 against Vapi's $0.00521 floor, 4.2% above. The pass-through claim holds. The doubling lives in a model-intelligence doc, not on the pricing page.

The model line is an estimate, not a rate. Vapi builds it from your system prompt and tool definitions, then assumes five model requests per turn, 150 output tokens a minute and a 50% cache hit rate. Change the prompt and the number moves, and Vapi warns that long calls usually cost more because conversation history grows.

The calculator offers three transcribers, three model providers and three voices. The platform supports many more, including xAI speech and voice added in June 2026, and none appear. Neither does a named model. Where Retell publishes a rate for each of 26 models, Vapi gives a range per provider.

How credits work

Credits are prepaid US dollars held by the subscription, and hosting and provider costs draw them down in real time.

Nothing in the docs says whether credits expire or can be refunded. Retell states both outright. Vapi states neither, which is the biggest gap on the page.

Purchase minimums go unpublished too. Buying credits needs "a minimum purchase" and auto-reload takes "at least the minimum amount", with no figure on either.

Free allowances are specific: $5 of credits, which at $0.05 hosting is exactly 100 minutes, matching the 100 minutes Vapi's startup page quotes. Spend it on a complete agent at $0.0681 and it lasts about 73 minutes. The free Vapi number is US only and inbound only, so an outbound agent can't use the $0 transport line and pays Telnyx or Twilio instead.

What happens when you hit the limit

Run out of credits and Vapi freezes the subscription. No new calls and no billable usage until you buy more, unless auto-reload fires on a threshold you set. That's a hard stop rather than an overage, so a production agent with auto-reload off goes dark.

Run out of concurrency and calls simply don't start. Usage only includes four lines, Core ten and Pro thirty. Past that, no outbound call starts and no inbound call connects until a slot frees. Vapi documents no queue, no wait and no fallback number, so what the caller hears is left to you.