Most production Voice AI calls land somewhere between $0.07 and $0.30 per connected minute at public self-serve prices. A five-minute call can therefore cost roughly $0.35 to $1.50 before implementation, support, and enterprise commitments. That range is wide because providers sell different units under similar “per-minute” language.
Vapi publishes a platform fee and passes model costs through. Retell publishes voice infrastructure, TTS, LLM, add-on, and telephony components. Bland publishes a bundled rate that includes STT, LLM, and TTS. Pipecat and LiveKit are closer to infrastructure layers, leaving model selection and carrier design in your hands. Comparing only the headline rate makes bundled platforms look expensive and component platforms look artificially cheap.
1. Platform and orchestration
The platform layer keeps a realtime agent running, coordinates turns, invokes tools, and manages the live session. Public pricing can be a pure usage fee, a monthly plan with included minutes, a voice-infrastructure rate, or a bundled agent minute. First classify the pricing structure, then decide which remaining components must be added.
Included concurrency also matters. A low per-minute fee can be overshadowed by paid lines when traffic bursts. Vapi's public Build plan includes a base number of concurrent calls and charges for extra lines. Retell includes a larger self-serve concurrency allowance and sells additional capacity. Bland ties concurrency to plan tiers. Infrastructure products make you responsible for the capacity model directly.
2. Speech to text
Streaming STT turns caller audio into tokens or text. It is commonly billed per audio minute or hour, sometimes with different rates for multilingual, medical, diarization, or conversational models. The price difference between efficient streaming models is often small compared with TTS or a large LLM, but transcription quality can affect every later stage of the call.
Silence handling and billing granularity matter. Per-second billing produces a different invoice from rounding each call or segment upward. Multichannel audio may also multiply billable duration. Use your provider's exact metering definition when moving from an estimate to a budget.
3. Language model inference
LLM cost is usually token-based even when a platform displays a convenient per-minute estimate. System prompts, tool definitions, conversation history, caller verbosity, agent response length, caching, and reasoning tiers all change the token count. A short appointment reminder can use a small, fast model; a regulated support call with long context and complex tools may need a more capable and more expensive model.
Use representative transcripts to estimate input and output tokens per turn. Do not assume the first turn's token use remains flat: conversation history can accumulate, particularly in speech-to-speech systems that repeatedly process audio context.
4. Text to speech and realtime audio models
TTS may be billed by character, token, credit, or generated minute. Voice quality, multilingual support, latency, and cloning features can create large price differences. An inexpensive platform paired with a premium voice can cost more than a bundled platform. Estimate the share of each call during which the agent speaks rather than charging TTS for the full connected duration.
Speech-to-speech models combine listening, reasoning, and generation into one realtime model. Their bills are not directly comparable with a traditional STT → LLM → TTS pipeline unless both are normalized to the same call length, turn pattern, and context behavior.
5. Telephony and transport
Phone calls introduce carrier charges. Inbound and outbound rates differ; local, toll-free, mobile, international, and SIP traffic differ again. Phone-number rental, transfers, media streams, call recording, answering-machine detection, branded calling, and failed-attempt minimums can each add a separate line item.
Model phone spend independently with thetelephony calculator. This is especially important for outbound campaigns, where failed or short calls can create a materially different cost per successful conversation.
6. Recording, analytics, and compliance
Recording commonly separates processing from storage. Transcription, summarization, quality assurance, PII removal, knowledge bases, denoising, and observability can all have their own per-minute or monthly charge. Retention policy affects storage costs and compliance exposure. For a reliable forecast, define what is recorded, how long it is kept, who can access it, and which post-call jobs run.
7. Concurrency and idle capacity
Monthly minutes describe volume, not arrival shape. Ten thousand minutes spread across a month can require one line or dozens, depending on business hours, campaign timing, and peak-to-average ratio. Managed platforms may charge per additional concurrent call. Self-hosted systems must provision enough idle compute to accept the burst.
Use the concurrency planner, then add headroom for transfers, retries, regional failover, deployments, and provider rate limits.
8. Managed cloud versus self-hosting
Self-hosting removes or reduces a platform fee, but creates an operating system around the agent. Include compute, networking, load balancing, autoscaling, realtime media tuning, observability, deployment safety, patching, incident response, and engineering time. Idle capacity is part of the price because a call cannot wait for a cold cluster during a burst.
Managed cloud is usually easier to justify during product discovery and early production. Self-hosting becomes financially interesting when traffic is sustained, the team already operates realtime infrastructure, or control and data residency are worth more than the direct platform savings.
A practical forecasting workflow
- Measure connected minutes, average handle time, direction, and success rate.
- Choose a platform architecture and classify bundled versus separate components.
- Price STT, LLM, TTS, carrier, recording, and add-ons at the same usage profile.
- Model peak concurrency separately from monthly volume.
- Add operational and compliance costs for a true ownership comparison.
- Request provider quotes using the same scenario and compare them with the public model.
Start with the Voice AI cost calculator, then replace its defaults with your own traffic and provider terms. A transparent estimate is not an invoice, but it is a much stronger purchasing baseline than one advertised rate.