OpenAI's Aarti said GPT-Live 1 delegates more reliably with a special SpawnThinking token.
She framed GPT-Live as full duplex talk plus delegation to a backend model.
Turn-based Realtime had to guess when you finished, so mm-hmms triggered false takeovers and noisy rooms broke it.
For intake needing name, date, time and address, Aarti said to send commentary for the next question and thinking updates that list the full order and block redelegation until confirmation.
Prompting can tell it to ignore side chat.
Aarti played Brise and Sillage, two new French voices, and said custom voices can come from prompt to voice or consented recordings.
Yelp uses it for reservations, Fin for support, Speak for lessons and Cognition for Devin voice.
OpenAI notes in release notes voice use in ChatGPT nearly doubled since GPT-Live launched, so keep frontend prompts short, save key facts outside auto compaction and use Decisions triggers to force missed delegation.
