Cost analysis and use case recommendations

You know the providers and their tradeoffs. Now it is time to put numbers to decisions. This chapter models the real cost of voice AI conversations across different stacks, then gives you concrete recommendations for four common use cases — each with exact LiveKit configurations you can copy into your project. All examples use LiveKit Inference where possible for the simplest setup, with plugin alternatives noted.

What you will learn

• How costs break down across STT, LLM, and TTS for a typical 5-minute conversation
• How to compare total cost across four stack profiles: Budget, Balanced, Premium, and Enterprise
• When realtime mode is cheaper or more expensive than pipeline mode
• Specific stack recommendations with LiveKit plugin code for common use cases
• Cost optimization strategies that do not sacrifice quality

Anatomy of a voice AI conversation cost

Here is a typical 5-minute customer service conversation broken into billable units:

Four stack profiles

Budget: Deepgram + GPT-4.1 mini + Cartesia

The lowest-cost pipeline that still delivers a good experience. Ideal for high-volume, straightforward interactions. All three models are available via LiveKit Inference.

from livekit.agents import AgentSession, Agent, RoomInputOptions, inference

# Via LiveKit Inference — no per-provider API keys needed
session = AgentSession(
  stt=inference.STT(model="deepgram/nova-3", language="en"),
  llm=inference.LLM(model="openai/gpt-4.1-mini"),
  tts=inference.TTS(model="cartesia/sonic-3", voice="9626c31c-bec5-4cca-baa8-f8ba9e84c8bc"),
)

await session.start(
  agent=Agent(instructions="You are a helpful customer service agent. Be concise."),
  room=ctx.room,
  room_input_options=RoomInputOptions(),
)

Balanced: Deepgram + GPT-4.1 + Cartesia

Strong reasoning with the fastest TTS. Good for agents that need reliable function calling and tool use. All available via LiveKit Inference.

Premium: Deepgram + Claude Sonnet + ElevenLabs

Best reasoning with the highest voice quality. Ideal for sensitive domains and brand-critical experiences. STT and TTS are available via Inference; Claude requires the Anthropic plugin.

from livekit.agents import AgentSession, Agent, RoomInputOptions, inference
from livekit.plugins import anthropic

# Mix Inference (STT, TTS) with plugin (LLM) as needed
session = AgentSession(
  stt=inference.STT(model="deepgram/nova-3", language="en"),
  llm=anthropic.LLM(model="claude-sonnet-4-20250514"),
  tts=inference.TTS(model="elevenlabs/eleven_flash_v2_5", voice="your-voice-id"),
)

await session.start(
  agent=Agent(instructions="You are a medical office assistant. Never provide medical advice."),
  room=ctx.room,
  room_input_options=RoomInputOptions(),
)

Enterprise: Azure STT + GPT-4.1 + Azure TTS

Compliance-first stack for regulated industries. HIPAA, SOC 2, and GDPR certifications across STT and TTS. Azure STT/TTS require the Azure plugin for private endpoints and compliance features; the LLM can use Inference.

from livekit.agents import AgentSession, Agent, RoomInputOptions, inference
from livekit.plugins import azure

# Azure plugin for compliance; Inference for LLM
session = AgentSession(
  stt=azure.STT(speech_key="your-key", speech_region="eastus", language="en-US"),
  llm=inference.LLM(model="openai/gpt-4.1"),
  tts=azure.TTS(speech_key="your-key", speech_region="eastus", voice="en-US-JennyNeural"),
)

await session.start(
  agent=Agent(instructions="You are a compliant enterprise assistant."),
  room=ctx.room,
  room_input_options=RoomInputOptions(),
)

Cost per minute summary

Pipeline vs realtime: cost comparison

Realtime models (OpenAI Realtime API, Gemini Live) charge per minute of audio rather than per token. Here is how they compare:

Cost optimization strategies

Use case recommendations

Customer service

High volume, straightforward interactions. Latency and reliability matter most. Fully available via LiveKit Inference.

Expected: 300-500ms latency, ~$0.03 per 5-min call.

Healthcare

Sensitive information, accuracy and safety paramount, compliance required.

Expected: 400-700ms latency, ~$0.07 per 5-min call.

Education

Patient tutoring, clear explanations, engaging voice, long sessions.

Expected: 400-700ms latency, ~$0.07 per 5-min session.

Entertainment

Character voices, personality, emotional range, immersion.

Expected: 400-700ms latency, ~$0.07 per 5-min session.

Quick reference: all stacks

What you learned

• A typical 5-minute voice AI call costs $0.02-0.07 depending on your stack
• The Budget stack (Deepgram + GPT-4.1 mini + Cartesia via LiveKit Inference) delivers the best cost-to-quality ratio for most use cases
• LiveKit Inference simplifies cost management with a single billing relationship across all providers
• Realtime mode (OpenAI Realtime API) costs 5-10x more than pipeline mode
• System prompt length, model tiering, and response conciseness are the biggest cost levers
• Choose your stack based on your top priority: cost, quality, compliance, or latency