Vapi

Vapi is the plumbing layer for voice AI. It does not try to be a finished phone-answering product; it is the developer platform you assemble one on top of. The job it does is orchestration — stitching together a speech-to-text engine, a large language model, a text-to-speech voice, and a telephony provider into a single low-latency loop that can hold a real-time conversation. For technical teams building custom voice agents — support lines, outbound qualification, appointment booking, IVR replacements — that orchestration is the hard part, and Vapi has become one of the cleanest ways to get it done in 2026.

The defining design choice is that Vapi is model-agnostic and bills by pass-through. You bring your own components: an LLM from OpenAI, Anthropic, or Groq; an STT engine; a TTS voice from a provider like Cartesia or ElevenLabs; and a telephony layer. Vapi coordinates them and charges for the coordination, while passing the underlying provider costs through to you at cost, without markup. That transparency is genuinely unusual and genuinely useful — you can see exactly what each layer of the conversation costs and swap any component for a cheaper or better one without leaving the platform.

Pricing starts simple and gets layered. The Vapi orchestration fee begins at about $0.05 per minute, with no subscription and no per-seat charge — you pay for what you run. But the orchestration fee is only the first of at least four cost layers: on top of it sit the LLM tokens, the STT minutes, the TTS characters, and the telephony charges, each passed through from its provider. In realistic configurations the all-in cost lands somewhere between $0.07 and $0.25 per minute, with simple setups using low-cost STT and TTS sitting near the bottom of that range. This is the single most important thing to understand before building on Vapi: the headline $0.05 is the platform’s slice, not your invoice.

At scale the numbers compound in ways worth planning for. Vapi includes 10 concurrent call lines by default, with additional lines at roughly $10 per line per month — a real cost for any operation expecting simultaneous calls. Typical moderate enterprise deployments land around $3,000-6,000 a month once volume and provider costs are included, and larger platform-access commitments can run $40,000-70,000 a year. None of that is hidden or unreasonable for the value, but it means Vapi rewards a team that actually models its expected concurrency and call volume before going live, rather than discovering the math on the first invoice.

The right comparison here is subtle, because Vapi sits at a different layer than some tools PixlRun covers. Cartesia Sonic 3, which we reviewed recently, is a text-to-speech engine — one of the components you would plug into Vapi, not a competitor to it. The actual competitors are other orchestration platforms such as Retell AI, which trade some of Vapi’s flexibility for a more guided setup. Against a packaged, no-code voice product, Vapi is the opposite philosophy: maximum control and component freedom in exchange for you doing the engineering. If you want to log in and click a voice agent into existence, Vapi is the wrong tool; if you want to build exactly the agent you need on the models you choose, it is among the best foundations available.

That positioning is also the main limitation. Vapi is an API platform aimed at developers and technical teams, full stop. There is no meaningful path for a non-engineer to deploy a production agent on it alone — you need someone who can write code, manage providers, and handle the telephony details. For a small business that just wants its phone answered, a turnkey receptionist product will be faster and cheaper to stand up. Vapi’s audience is the team building that product, or building something more bespoke than any off-the-shelf option allows.

The verdict: for engineers building real-time voice agents, Vapi is one of the most flexible and honestly-priced platforms on the market in 2026. The model-agnostic, pass-through approach means you are never locked into one vendor’s voice or one vendor’s LLM, and you can optimise cost and quality component by component. The two things keeping it from a higher score are inherent to its nature: the multi-layer pricing genuinely takes effort to forecast, and the developer-only ceiling puts it out of reach for the non-technical. Build voice agents for a living and Vapi belongs on your shortlist; want your phone answered tomorrow and you want something else.

The developer experience is a genuine part of the value. Vapi ships SDKs, clear documentation, and a dashboard for configuring assistants, defining conversation flows, and inspecting call transcripts and recordings after the fact. That observability matters more than it sounds: debugging a voice agent means listening to where the conversation broke, and Vapi gives you the transcript and the component-level latency to find it. Function calling lets the agent trigger real actions mid-call — looking up an order, booking a slot, escalating to a human — which is what separates a useful agent from a glorified phone tree. For a team that already has engineers, the on-ramp is days, not weeks, and the platform gets out of the way once the pipeline is wired.