A voice AI agent running 10,000 minutes/month on a managed platform costs $2,300–$3,300/mo in infrastructure. Add $200–$500/mo in platform and ops. Build cost: $10K–$40K one-time. Total first-year cost for a production deployment at this volume: approximately $40K–$80K.
The 3 cost layers
Layer 1: Infrastructure (per-minute, ongoing)
Every voice AI call burns compute in real time across four components: speech-to-text (STT), a large language model (LLM) for response generation, text-to-speech (TTS) for audio output, and a telephony layer for PSTN call handling. These four costs combined run $0.07–$0.33/min depending on platform and scale.
The platform choice determines most of this cost. Managed platforms like Vapi and Retell handle infrastructure for you at $0.21–$0.33/min. Self-hosting your own LiveKit stack at scale drops to $0.07–$0.15/min — but adds significant engineering overhead.
Layer 2: Build cost (one-time)
A voice AI agent isn't a pre-trained product you subscribe to — it needs custom conversation design, integration with your backend systems (CRM, calendar, ticketing), error handling for edge cases, and production monitoring. That engineering work is a one-time investment that doesn't repeat.
- Proof of Concept (production-ready demo on your real use case): $5K–$15K, 2–4 weeks
- Full production build with CRM and telephony integration: $10K–$40K
- Enterprise builds with compliance, multi-language, complex orchestration: $50K–$100K+
Layer 3: Ongoing operations (monthly)
Platform subscriptions, telephony number fees, monitoring tooling, and quarterly prompt tuning as your business changes. Typically $200–$2K/mo depending on volume and how much the conversation flows evolve over time.
Per-minute cost by platform
| Platform | Per-minute cost | Best for | Build complexity |
|---|---|---|---|
| Vapi (BYOK) | $0.23–$0.33/min | API-first, outbound SDR, fast deployment | Low |
| Retell AI (BYOK) | $0.21–$0.30/min | HIPAA, regulated industries, visual flow editor | Low |
| LiveKit + OpenAI RealtimeSWEET SPOT | $0.15–$0.20/min | Sub-300ms latency, 10K+ min/mo | Medium |
| LiveKit (self-hosted) | $0.07–$0.15/min | 50K+ min/mo, data sovereignty | High |
Total infrastructure cost by monthly volume
| Monthly minutes | Vapi / Retell | LiveKit + Realtime | LiveKit (self-hosted) |
|---|---|---|---|
| 1,000 min/mo | $230–$330/mo | $150–$200/mo | $90–$150/mo |
| 10,000 min/mo | $2,300–$3,300/mo | $1,500–$2,000/mo | $700–$1,500/mo |
| 50,000 min/mo | $11,500–$16,500/mo | $7,500–$10,000/mo | $3,500–$7,500/mo |
| 100,000 min/mo | $23,000–$33,000/mo | $15,000–$20,000/mo | $7,000–$15,000/mo |
The self-hosted crossover point is around 50K minutes/month. Below that, the DevOps and infrastructure management overhead of self-hosting erodes the savings. Above it, the 50–60% infrastructure reduction is meaningful enough to justify a dedicated ops budget.
What drives cost up or down
- Using GPT-4o instead of GPT-4o-mini (3–5× higher LLM cost for similar tasks)
- Long average call durations — every minute of silence still burns infra
- Premium TTS voices (ElevenLabs cloned voices vs standard Azure or OpenAI TTS)
- High concurrency without pre-provisioned capacity — causes queuing and wasted infra
- Low call volume — fixed platform fees spread over fewer minutes raises effective per-minute cost
- BYOK (bring your own API keys) on Vapi or Retell cuts platform markup
- GPT-4o-mini for most turns, GPT-4o only for complex reasoning branches
- Self-hosting on LiveKit at 50K+ minutes/month — 50–60% infra cost reduction
- Good conversation design: shorter conversations with higher first-call resolution
- Batching non-real-time work (logging, CRM sync) outside of the live call window
Frequently asked questions
How much does a voice AI agent cost per minute?
Between $0.07 and $0.33 per minute. The range depends on platform and volume. Managed platforms (Vapi, Retell) with BYOK pricing run $0.21–$0.33/min. Self-hosted LiveKit infrastructure at 50K+ minutes/month drops to $0.07–$0.15/min. For most deployments under 20K minutes/month, managed platforms are cheaper when you factor in the DevOps cost of self-hosting.
What is the total build cost for a voice AI agent?
A production voice AI agent with custom conversation flows, CRM integration, error handling, and monitoring runs $10K–$40K. A Proof of Concept validating latency, accuracy, and user experience on your real use case: $5K–$15K in 2–4 weeks. Enterprise builds with compliance requirements, multi-language support, and complex orchestration run $50K–$100K+.
Is Vapi or Retell cheaper for voice AI?
They're within $0.02–$0.03/min of each other at comparable configurations. Platform choice should be driven by requirements, not marginal cost differences: HIPAA compliance points to Retell, visual conversation flow editor points to Retell, API-first and outbound SDR workflows point to Vapi. See our full Vapi vs Retell comparison for the detailed breakdown.
At what call volume does self-hosting voice AI become cost-effective?
Around 50,000 minutes/month. Below that threshold, managed platforms (Vapi, Retell) are cheaper once you factor in Kubernetes cluster management, monitoring infrastructure, and DevOps time. Self-hosting LiveKit at 50K+ min/mo cuts infrastructure costs by 50–60%, but the ops overhead requires a dedicated engineering resource.
What drives voice AI agent cost up?
The biggest factors: (1) LLM model choice — GPT-4o costs 3–5× more than GPT-4o-mini for equivalent tasks. (2) Long average call duration — every minute of silence still burns infrastructure. (3) Premium TTS voices like ElevenLabs cloning vs standard OpenAI or Azure voices. (4) Low monthly volume — fixed platform fees dominate at small scale.
How much does voice AI cost at enterprise scale (100K min/month)?
On LiveKit self-hosted: $7K–$15K/mo in infrastructure costs. Add $1K–$3K/mo for monitoring, DevOps, and prompt maintenance. Total: $8K–$18K/mo for 100,000 minutes. Compare that to a human call center of 5–10 agents handling the same volume: $25K–$60K/mo in salaries alone, plus training, benefits, and management overhead.