OpenAI

One OpenAI key covers everything: text replies, realtime voice, and both halves of the cascade.

Key: create one at platform.openai.com/api-keys.

What it's used for

Use	Supported
Text replies (the brain)	Yes: GPT models and reasoning models
Images and camera frames	Yes, except gpt-3.5, o1-mini and o3-mini (the admin panel warns)
Tools	Yes
Realtime voice	Yes: the Realtime API (voice.mode: realtime with an OpenAI connection)
Speech to text (cascade)	Yes: realtime transcription, gpt-4o-mini-transcribe by default
Text to speech (cascade)	Yes: gpt-4o-mini-tts by default

Models

The server reads your key's model list and sorts it by use. New bots start with the newest gpt-N model for text and gpt-realtime for realtime voice (when your key has them).

• Reasoning models get brain.effort as their reasoning effort, and keep their reasoning from one turn to the next. brain.temperature only applies to non-reasoning models.

Realtime voice

• Turn detection (voice.realtime.turnDetection): semantic_vad (the default) waits until the visitor seems to have finished their thought, with eagerness from low to high (or auto); server_vad answers after about half a second of silence.
• Voices: marin and cedar (recommended), alloy, ash, ballad, coral, echo, sage, shimmer and verse. Preview them from the admin panel.
• Camera: sees frames with each turn, or when the agent uses look_at_camera. It doesn't take a live video stream (use Gemini Live for that).
• Interrupted replies are cut where the visitor stopped hearing them, so the model's memory matches what was said.

Text to speech

All the realtime voices plus fable, nova and onyx. voice.cascade.tts.speed sets the speed, and voice.cascade.tts.style is passed as instructions ("warm and unhurried").