Vapi Review (2026)

Vapi charges $0.05 a minute and that number is honest, as far as it goes. It buys orchestration: turn detection, interruption handling, the state machine that decides when your agent should stop talking, and the plumbing between four services you are expected to bring yourself. What it does not buy is any of those four services.

This is the bring-your-own-key model, and the category has largely stopped explaining it. A buyer reading the pricing page sees a rate three to four times below the managed platforms and reasonably concludes it is three to four times cheaper. It is not. Once speech-to-text, the model, synthesis and telephony are added at published list rates, the floor lands at $0.0965 — about 1.9× the headline, and squarely in the middle of the category rather than below it.

The vendor's own calculator answers the question

Vapi publishes a usage calculator, and nobody writing about Vapi's pricing appears to have driven it to a total. We did, on 2026-08-06.

Left on its own defaults, a thousand calls of a minute each, and configured with the three model providers Vapi itself names in its FAQ — Deepgram for speech, OpenAI for the model, ElevenLabs for synthesis — plus Twilio for the line, it returns a total of $110.30.

The calculator also answers how much this varies. Choose the cheapest selectable option on every menu and the stack lands near $0.055 a minute. Choose the dearest and it passes $0.32 before the model is counted. A 5.9-fold spread, all of it inside what people call Vapi's price.

Worth noticing while you are in there: Vapi appears in its own synthesis menu at $0.0216 a minute, selling its own voice beside the third parties, on a page that frames model cost as somebody else's pass-through.

Where the money actually goes

The four excluded layers are not equal. Synthesis dominates, because the agent speaks roughly half of a two-party call and character-based pricing is unforgiving at conversational speed. The model comes second and grows through the call, since every turn re-sends an ever-longer transcript. Recognition and telephony are close to rounding errors by comparison, which is the opposite of what most buyers assume.

The practical consequence is not that any one line dominates, it is that only one of them has real range. Telephony is a carrier rate you cannot argue with, recognition varies by tenths of a cent, and the model is 11% of the bill however hard you tune it. Synthesis is the exception: the same minute of call costs $0.014 on Cartesia and around $0.085 on premium ElevenLabs. That single choice swings the total further than every other component combined, and on a phone call band-limited to 8kHz the difference is frequently inaudible.

What you are taking on

Four vendor relationships, four sets of rate limits, four status pages and four invoices. When a call fails you own the diagnosis, and the failure is often at a seam rather than inside any one component. Interruption handling in particular is a tuning problem that does not have a correct default — endpointing thresholds that feel responsive on a quiet line clip people on a mobile in a car park.

That is the actual cost comparison, and it is not on anybody's pricing page. Against a managed platform at $0.07 bundled, Vapi is cheaper per minute at scale and more expensive for the first three months, because the first three months are engineering.

Who should not buy this

If nobody on your team wants to reason about endpointing thresholds, voice model selection, or which of four vendors caused last night's dropped calls, a bundled platform will cost you less in total. That is not a knock on the product, it is the correct reading of a product that deliberately hands you the controls.

Self-serve, free credit, and a working agent inside an hour if you already hold accounts with a speech, model and telephony vendor. If you do not, budget a day for account setup before you write any agent logic.

The build is not the hard part. The tuning is, and it is where the timeline actually goes.

• Hold four accounts before starting. Speech, model, synthesis and telephony. Each has its own rate limits, its own status page and its own billing cycle.
• Tune endpointing on a real phone line, not on a laptop microphone. Thresholds that feel responsive in an office clip people speaking from a car.
• Test barge-in deliberately. Talk over your own agent mid-sentence and see whether it stops. This is the failure that makes a deployment feel broken rather than slow.
• Budget for the seams. When a call fails you own the diagnosis, and the fault is frequently between components rather than inside any one of them.

Published and self-serve at $0.05 a minute for orchestration, with every model layer billed by whoever supplies it.

The all-in floor computed on published component rates is $0.1217, which is 2.4x the headline. That figure assumes mid-market defaults. The point of the architecture is that you can beat it.

Every row below the first requires someone to have made deliberate choices and to revisit them as the market moves. That is the actual product decision here, and it is a staffing question rather than a pricing one.