OpenAI appears to be preparing a new voice model for ChatGPT, GPT-Bidi-1, that can listen and speak simultaneously rather than waiting for the user to finish before responding. The company has not announced the model, but references to it have appeared in ChatGPT’s web and mobile code, according to TestingCatalog and Android Authority.
Why bidirectional matters
Today’s voice assistants, including ChatGPT’s existing voice mode, work mostly in turns: the user speaks, the system processes the audio and then replies. Interrupting usually means stopping the response and starting again. A bidirectional, or full-duplex, model keeps listening while it talks, which is closer to how people actually converse.
In practice, that would let a user cut in mid-answer, ask a follow-up or change the subject without the assistant losing its place. Early testing described in the reports suggests the model handles abrupt topic shifts, interruptions and fast back-and-forth exchanges more smoothly than turn-based systems.
What has surfaced so far
TestingCatalog, which tracks unreleased features in consumer apps, says GPT-Bidi-1 first appeared in ChatGPT’s code around June 15. The site subsequently reported that OpenAI was preparing a refreshed voice interface on the web, distinguished by a new yellow color scheme, to accompany the rollout.
Multiple references across the web and mobile apps have led observers to conclude that a release to users is close, although OpenAI has not confirmed a date, pricing tier or which users would get access first. As with any feature spotted in code before launch, the details could change or the rollout could be delayed.
A shift toward natural conversation
If the feature arrives as described, conversations with ChatGPT would feel less like issuing commands and more like talking with a person who can be interrupted and redirected. The development fits a broader pattern across the industry, where companies are competing not only on the quality of answers but on how natural the interaction feels, with low latency and real-time responsiveness becoming selling points in their own right.
OpenAI has previously invested heavily in voice, including the real-time audio capabilities introduced with its GPT-4o models. GPT-Bidi-1 would extend that work by removing the turn-taking constraint that still defines most voice assistants.
