If you've been exploring Vapi AI for voice agent infrastructure, you've probably noticed the pricing page raises more questions than it answers. Vapi's self-serve tier, Build, is usage-based, which sounds simple until you realize your invoice stacks at least four cost layers on top of each other.
This article breaks down Vapi AI pricing in full: what each layer covers, how they interact, and what you'll realistically pay. For teams needing phone-based scheduling, we also compare Zeeg — Europe's leading AI phone assistant and appointment scheduling software in one solution, GDPR-compliant and made in Germany.
Zeeg's AI agent answers calls, qualifies leads, and books meetings — no providers to configure, no code to write. Try any paid plan free for 14 days.
Sign up for freeWhat is Vapi AI?
Before digging into the numbers, here's a quick orientation for you. Vapi is a developer-focused platform that lets engineering teams build, deploy, and scale voice AI agents. It handles the infrastructure layer (real-time audio processing, latency management, and the orchestration between different AI model providers) so developers don't have to build that themselves.
It's mainly used for things like inbound customer service agents, outbound sales dialers, appointment reminder bots, and similar phone-based automation. The platform connects to external providers for speech-to-text (STT), text-to-speech (TTS), large language models (LLM), and telephony. And here's the thing, each of those providers charges its own fee. Vapi's cost is the sum of all of them, plus its own hosting fee on top.
That layered structure is precisely why Vapi AI cost can feel hard to pin down. So let's go layer by layer.
The four cost layers in Vapi AI pricing
There's no single "Vapi price" as what you pay is the combined total of four distinct components. Most pricing pages for tools like this present a tidy number, but Vapi is unusually transparent about what goes into it. That's both refreshing and, admittedly, a bit daunting at first.
Here's what you're actually paying for:
1. Vapi hosting cost — this is the platform fee Vapi charges for running your agents. As of 2026, it's $0.05 per minute for calls.¹ With no subscription or seat fees.
2. Large Language Model (LLM) costs — this is what you pay to the AI model processing the conversation. You can use OpenAI, Anthropic, Groq, or others. These are charged at cost from the provider, Vapi passes through whatever the LLM provider charges, without markup.
3. Text-to-Speech (TTS) costs — the voice synthesis layer. ElevenLabs, Deepgram, OpenAI, and others are supported. Again, at-cost pricing from the provider.
4. Speech-to-Text (STT) costs — transcription of the caller's voice. Deepgram is the default, though others are available. Same deal: at-cost passthrough.
Telephony (the actual phone call transport) is a fifth consideration, though it's billed separately through your own telephony provider like Twilio, Vonage, Telnyx, or similar. Vapi doesn't resell phone numbers or minutes directly; you bring your own provider.
Breaking down Vapi hosting cost
The $0.05/minute Vapi hosting fee is the fixed part of the equation. It covers the infrastructure for running your agent: real-time audio processing, the WebSocket connections, the orchestration between providers, and the platform itself.
For context: 1,000 calls at an average of 5 minutes each equals 5,000 minutes, which is $250 in Vapi hosting alone before adding any model provider costs. That number scales linearly with usage, which is predictable but worth factoring in from the start.
Vapi also charges for call concurrency. The Build plan includes 10 concurrent lines, with additional lines available at $10 per line per month.² If you're running a high-volume outbound campaign where dozens of calls happen simultaneously, that cost adds up.
For SMS and chat interactions, pricing is $0.005 per message which is a lot less than voice, though it depends on your use case.
LLM provider costs: The wildcard
This is where the numbers start to diverge significantly depending on your setup. LLM costs are billed at cost from whichever provider you use, and the per-minute equivalent varies a lot.
A few reference points at current published rates:
- GPT-5.6 Luna (OpenAI): $0.20 per 1M input tokens and $1.20 per 1M output tokens — the cheapest current OpenAI tier, and the sensible default for straightforward call handling³
- GPT-5.6 Terra (OpenAI): $2.00 input / $12.00 output per 1M tokens, for higher-volume work that needs more reasoning
- GPT-5.6 Sol (OpenAI): $5.00 input / $30.00 output per 1M tokens — the flagship, and roughly twenty times Luna's cost per conversation
- Claude Haiku 4.5 (Anthropic): $1 input / $5 output per 1M tokens — about four times Luna's rate, still well below the flagship tier⁴
- Groq-hosted Llama models: among the cheapest options, particularly for latency-sensitive voice use cases
Worth noting if you are reading an older cost breakdown: GPT-4o and GPT-4o mini are no longer part of OpenAI's current pricing lineup and survive only as legacy model IDs. Any per-minute maths built on them is a generation out of date.
An average one-minute phone call might involve somewhere between 500 and 2,000 tokens depending on conversational complexity, the length of your system prompt, and how verbose the caller is. At those volumes, the model layer costs roughly $0.001 per minute on Luna and around $0.02 per minute on Sol — a twentyfold spread driven entirely by model choice rather than by call volume.
If you bring your own API key for the LLM provider, Vapi still charges its $0.05/min hosting fee, but the LLM costs go directly to your provider account rather than through Vapi. That means you get your provider's native pricing without any additional margin.
TTS and STT costs
One of these layers is small. The other is the one most likely to dominate your bill, and it is not the one people expect.
Speech-to-Text (STT): Deepgram is the default option on Vapi. Note that Nova-2 has been superseded by Nova-3, which is what Deepgram now prices on its public rate card. Nova-3 monolingual streaming lists at $0.0077 per minute at standard rates, currently discounted to $0.0048; the pre-recorded rate is $0.0043 per minute.⁵ Streaming is the relevant number for live calls, so budget against that rather than the pre-recorded rate. AssemblyAI and others are also supported at similar price points.
Text-to-Speech (TTS): This is the layer that varies most, and the one worth checking carefully, because providers do not even bill in the same unit. ElevenLabs prices per minute rather than per character: extra minutes run about $0.36 on the free tier and settle at roughly $0.17 on paid plans, with low-latency TTS advertised as low as 5 cents per minute on the Business plan.⁶ Deepgram's Aura-2 bills per character instead, at $0.030 per 1,000 characters — on the order of $0.013 per minute at typical agent speaking rates, which makes it roughly a tenth the cost of ElevenLabs on a mid-tier plan. OpenAI no longer publishes a standalone TTS rate at all; its current voice pricing is folded into the GPT-Realtime models and billed per audio token.
The voice you choose has a real cost implication, in other words — and at the premium end it can cost more than every other layer of your stack combined.
Telephony costs: The layer Vapi doesn't cover
Here's something that catches people off guard: Vapi does not handle telephony directly. You need a separate telephony provider (Twilio, Vonage, Telnyx, Bandwidth, or similar) to actually place or receive calls. Vapi connects to your provider via SIP or PSTN, but the phone number rental and call minutes are billed by that provider independently.
Twilio, for example, charges:⁷
- $0.0085/minute to receive calls on a US local number
- $0.0140/minute to make calls to US local numbers
- $1.15/month per local US number, or $2.15/month for toll-free
Telnyx tends to be cheaper, with published SIP trunking rates starting at $0.0032/minute inbound and $0.005/minute outbound — one of the few places where outbound is not the more expensive direction.⁸ Vonage and Bandwidth have their own rate structures. The point is that this cost layer is real, it's not small at scale, and it's entirely separate from anything Vapi invoices you for.
What does a real call actually cost?
Let's put it all together with a realistic example. Assume:
- A 3-minute inbound customer support call
- GPT-5.6 Luna as the LLM
- Deepgram Nova-3 for STT
- Deepgram Aura-2 for voice
- Twilio for telephony
Here's a rough cost estimate per call:
The figures below combine published provider rates with our own usage assumptions. Token and character volumes vary with prompt length and conversation style, so treat the totals as a modelled estimate rather than a quoted price.
So for 1,000 calls at 3 minutes each, you're looking at roughly $230 per month in total per-call costs, using an economical model and a mid-cost voice.
Now swap GPT-5.6 Luna for Sol and Aura-2 for ElevenLabs on a paid plan, and the same 3-minute call runs closer to $0.75–$0.80 — more than triple. The striking part is where the increase comes from: the flagship model adds about seven cents, while the premium voice adds roughly half a dollar. At 10,000 calls a month that single choice is a $5,000 line item, which is why the voice, not the model, is usually the first thing to optimize.
Concurrency, data retention, and add-on costs
Beyond the per-minute charges, there are a few other cost factors worth knowing.
Concurrency is probably the most overlooked one. By default, Vapi gives you 10 concurrent lines. For a low-volume setup, that's fine. But if you're running outbound campaigns or handling peak inbound volumes, you may need more; at $10 per additional concurrent line per month. For a contact center running 50 simultaneous calls, that alone is an extra $400/month.
Data retention is also something to think about. On the Build plan, call history is kept for 14 days and chat history for 30 days. If your use case requires longer records (for compliance, for example) that's a paid add-on. HIPAA and Zero Data Retention are two separate add-ons, not one: HIPAA costs $2,000/month and Zero Data Retention $1,000/month, at the same price on both the Build and Scale tiers. A healthcare deployment that needs both is therefore paying $3,000/month before a single call is placed, which tells you something about the target market at that tier.
Scale-tier features like SSO, RBAC, SOC2 compliance, named support engineers, and SLAs sit on Vapi's Scale plan — a quoted annual contract with a fixed platform fee and committed volume — and Vapi doesn't publish those prices publicly. You'll need to contact their sales team, which is standard practice for enterprise tiers but worth flagging if you're planning a larger deployment.
Is Vapi AI pricing actually competitive?
Relative to building the same infrastructure yourself, yes; Vapi is genuinely cost-efficient. Handling real-time audio orchestration, managing the latency between multiple AI providers, and keeping everything stable at scale is not trivial engineering work. The $0.05/min hosting fee is a reasonable charge for avoiding that.
Compared to other voice AI platforms, Vapi sits on the lower end of the pricing spectrum for similar capabilities. Platforms like Retell AI, Bland AI, and others in the space tend to price in the $0.11–$0.31/min range for an all-in-one hosted solution. The trade-off is that those platforms abstract away the provider complexity as you pay a single fee and don't have to manage STT, TTS, and LLM configurations yourself.
Vapi's model gives you more control and potentially lower costs at scale if you optimize your provider stack. But that flexibility requires more engineering investment. If you're a solo operator or a small team without dedicated developers, the overhead of managing four separate provider accounts and bills might not be worth the savings.
It is worth putting a number on that overhead. The US Bureau of Labor Statistics puts the median wage for software developers, QA analysts and testers at $63.20 an hour — $131,450 a year for the occupation as a whole, and $133,080 for software developers specifically.⁹ A stack that takes even four hours a month to maintain across four provider accounts costs roughly $250 a year in engineering time before a single call is placed. At ten hours a month, the maintenance costs more than the hosting.
Who is Vapi actually built for?
Honestly, Vapi is a developer's tool. The documentation, the architecture, and the pricing model all point to a primary audience of engineering teams building voice AI products at scale. Companies using Vapi typically include:
- SaaS companies building voice features into their own products
- Agencies developing AI phone agents for clients
- Larger businesses with dedicated dev resources deploying call center automation
It's not really well-suited for small businesses or non-technical teams looking for a plug-and-play AI phone agent. There's no visual no-code builder, no out-of-the-box booking integration, and no CRM layer. You get the infrastructure, what you build on top of it is up to you.
That distinction has a cost attached. The role most voice agents are bought to support — the receptionist answering the phone — has a median wage of $17.90 an hour.¹⁰ The engineer required to assemble and maintain a Vapi stack costs $63.20. For a business whose goal is answering calls rather than shipping a voice product, the build is not competing against the platform fee; it is competing against the salary of the person who would otherwise pick up.
That's a meaningful distinction if you're evaluating tools for appointment booking, lead qualification, or customer scheduling specifically.
Zeeg: A different angle on voice AI for scheduling
Speaking of which, if your main goal is AI-powered phone scheduling rather than custom voice infrastructure, it's worth looking at Zeeg before committing to a Vapi build.
Zeeg is a scheduling CRM that has a natively built voice AI agent, meaning the calling, booking, and CRM layers are all in one platform. There are no separate accounts to manage across Twilio, OpenAI, ElevenLabs, and Vapi — it's all handled under one roof.
The difference in setup experience is significant. With Vapi, getting a working voice agent requires selecting and configuring multiple providers, writing prompts, setting up telephony, and building your own booking logic on top. With Zeeg, you write a prompt, pick a phone number, set routing rules, and your agent is live in a matter of minutes with no code required.
From a cost standpoint, Zeeg sells AI voice as monthly minute bundles rather than per seat.¹¹ The smallest bundle is €49/month billed annually, covering up to 250 inbound minutes — about €0.196 per inbound minute, or €0.274 per outbound minute at 179 minutes. Larger bundles bring the inbound rate down to €0.111/min. Every call logs its exact cost in the CRM, so there are no surprise invoices at the end of the month.
Here's how the two platforms compare for a scheduling-focused use case:
The two tools aren't really direct competitors as Vapi is infrastructure, while Zeeg is an end-to-end scheduling platform. But if your use case is getting an AI agent to answer calls and book meetings, Zeeg is worth evaluating seriously before choosing to build on Vapi from scratch.
Zeeg's voice AI agent handles inbound and outbound calls, routes based on caller intent using plain-language rules, captures lead details, and books appointments, all logged automatically in the CRM. You can import an existing SIP number or purchase a local number directly in the platform. The AI voice minute bundles are priced separately from the scheduling seats, so you can start with voice alone; scheduling plans begin at €10/user/month billed annually.
No Twilio account. No LLM setup. No separate booking tool. Zeeg handles it all, with transparent per-call costs logged directly in your CRM.
Try Zeeg freeWrapping up on Vapi AI's 2026 costs
Vapi AI pricing on the Build tier is genuinely usage-based, which is good for flexibility. But the true Vapi AI cost is the sum of at least four billing layers (platform hosting, LLM, TTS, STT) plus a separate telephony provider you manage yourself. For a simple 3-minute inbound call with an economical model stack, you're looking at roughly $0.23 per call. Move to a flagship model and a premium voice and the same call approaches $0.80 — and almost all of that increase is text-to-speech, not the language model.
It's a competitive model for engineering teams that want full control over their voice AI stack and are comfortable managing multiple provider accounts. For teams who just want AI-powered phone scheduling without the build overhead, a more integrated platform like Zeeg offers a comparable per-minute cost with a fraction of the setup complexity and the booking, CRM, and call records all come built in.
Pricing last verified: 9 September 2026. AI voice pricing changes frequently — figures reflect published rates on that date.
Sources
- Vapi pricing
- Vapi documentation
- OpenAI API pricing
- Claude Platform — API pricing
- Deepgram pricing
- ElevenLabs pricing
- Twilio Voice pricing
- Telnyx SIP trunking pricing
- US Bureau of Labor Statistics — Occupational Outlook Handbook: Software Developers, Quality Assurance Analysts, and Testers
- US Bureau of Labor Statistics — Occupational Outlook Handbook: Receptionists
- Zeeg pricing explained — AI voice minute bundles
