OpenAI appears to be rolling out a new bidirectional voice mode for ChatGPT, a feature that lets the assistant speak and listen at the same time. References to the new system have started showing up in the ChatGPT web interface, and early access has reportedly reached a limited group of users in the app.
The new capability is associated with a next-generation audio model identified as Bidi 1. The name appears to refer to the model’s bidirectional design, which allows ChatGPT to process input and output more fluidly during a conversation. In practice, that means the assistant can stay engaged while a user is talking, rather than waiting for a more rigid turn-based exchange.
Initial testing suggests the new mode behaves differently from ChatGPT’s current advanced voice experience. In the model selector, Bidi 1 appears alongside standard and advanced voice options, and the voice bubble changes color when it is selected. The model also gives short acknowledgments, such as brief verbal confirmations, when a user pauses or slows down. Importantly, it does this without interrupting the speaker.
The feature also seems better at handling changes mid-conversation. In one example from testing, ChatGPT was asked to count to ten and then interrupted with a request to reverse the count. The assistant adjusted on the fly, suggesting more responsive conversational control than the current voice stack provides.
Another notable improvement is continuity. The new voice mode appears to preserve the thread of a conversation more effectively instead of losing earlier context, which has been a common complaint about existing voice interactions. It also no longer appears to jump in during longer pauses, a behavior that can make some voice assistants feel overly eager or disruptive.
Some of the more playful behavior from the original advanced voice rollout is still present. According to the source material, the system can still sing and beatbox. But copyright handling appears more restrictive this time. It reportedly refuses requests for popular songs, while still trying to produce original material in the style of a chosen artist.
That shift suggests OpenAI is refining how the voice system balances creativity with safeguards. The company has not formally announced the rollout, so the details remain based on early observations rather than an official product briefing.
The feature seems likely to launch gradually across web and mobile in an opt-in format. The source material also suggests users in the European Economic Area may see the feature later, although that has not been confirmed.
Beyond ChatGPT, the company appears to be planning related voice work for Codex in the weeks after this launch, though that upgrade is expected to be separate. API access may follow later as well, but the timing has not been confirmed.
If the rollout continues as observed, Bidi 1 could mark one of the biggest changes to ChatGPT’s voice experience in months. By making conversation feel more continuous and less interrupt-driven, OpenAI seems to be pushing voice mode closer to a natural back-and-forth assistant rather than a simple speech interface.