Realtime Voice API Cost
Quick Answer
Realtime voice API cost can depend on audio input, audio output, speech-to-text usage, LLM tokens, text-to-speech output, session duration, latency settings, interruptions and retry behavior. Blending these into one session cost estimate requires understanding how the provider meters each component.
What affects cost
Realtime voice cost is shaped by STT token consumption, LLM reasoning tokens, TTS output, session length, audio bitrate, interruption handling that restarts the LLM turn, and whether the provider charges for active listening time or only output generation.
Common billing units
- audio for speech-to-text input
- token for LLM reasoning and context
- audio_out for text-to-speech output
- session for connection time or per-turn billing
Cost risks
Long sessions without output, interruption-heavy flows restarting LLM turns, high audio bitrate increasing STT cost, and assuming a simple token-only model can all make realtime voice cost higher than expected.
Small test checklist
- Test one short session and record all cost components
- Check STT vs. LLM vs. TTS breakdown if available
- Test interruption-heavy scenarios for LLM turn restart cost
- Compare session billing with output-only billing
Common errors / failed tasks
Session disconnect causing duplicate LLM turns, interruption without proper turn handling, audio format mismatch, high bitrate increasing STT cost and session timeout resubmission. Use logs to verify each component.
Related pages
OpenAI API Usage
Read the related cost or billing topic
API Billing Mismatch
Read the related cost or billing topic
Media API Cost Hub
Read the related cost or billing topic
Billing Transparency
Read the related cost or billing topic
Realtime voice API cost blends STT, LLM tokens, TTS output, session duration and interruption handling. This page is educational and helps developers planning realtime voice workflows understand the blended cost model. Check live provider pricing before production use and test small before scaling sessions.
Frequently Asked Questions
Is realtime voice billed like a chat API?
Not exactly. Realtime voice often blends STT, LLM and TTS into one session, each with potentially different billing units.
Do interruptions add to the bill?
They can, if each interruption restarts an LLM turn that is billed separately.
Should I test session cost separately from output cost?
Yes. Some providers charge for connection time while others only charge for output. Test both scenarios.
Start with a small prepaid test
Create an API key with $1 trial credit and test realtime voice session cost before scaling.