Every call's cost is broken down across three components, each with its own billing unit:
| Component | Billing Unit |
|---|
| Transcriber | Minutes of audio |
| Model (LLM) | Per 1 million input and output tokens |
| Voice | Characters of spoken text |
Transcriber cost scales with call duration, since audio is transcribed from both the caller and the assistant.
Voice cost increases when the assistant speaks more, since it's based on the number of characters spoken.
Model cost is the most variable component, since it depends on your system prompt length, tool definitions, and how much the conversation grows over the course of a call. Displayed costs are estimates for comparison, not exact quotes — large, well-cached prompts often cost less than estimated, while long calls often cost more, since the growing conversation history is re-sent with each request.
To reduce costs: shorten your system prompt and tool definitions, use models that support prompt caching, choose a lower-cost Model Preset when the use case allows, and keep calls focused and shorter.