Leaks point to OpenAI voice model 'Bidi 1' that sings and handles interruptions

An unreleased OpenAI voice model said to be in the works, dubbed "Bidi 1" (GPT-Bidi-1), reportedly can sing, generate a variety of sounds, and respond to interruptions mid-conversation. Sample clips shared on June 23, 2026 showcased a wide range of audio generation including singing and emotional expression, drawing attention among developers. It should be noted, however, that this is not an official announcement but leaked information based on in-app code references and limited testing.