OpenAI is testing an unannounced bidirectional voice model called Bidi 1 that lets ChatGPT listen and speak at the same time.
Key Points:
- Bidi 1 can listen, speak, and absorb mid-sentence interruptions without freezing the conversation.
- Code references surfaced in mid-June, and OpenAI has made no formal announcement.
- The model has begun reaching some app users, hinting at a release as soon as this week.
Bidi 1 Surfaces In ChatGPT Code
Code and interface elements tied to the model first appeared inside the ChatGPT app around Jun. 16, weeks before any formal reveal from a company that has said nothing publicly. The new option sits in the model selector under settings, resting beside the standard and advanced voice modes that users already know. Pick it, and the voice bubble glows yellow.
The name is shorthand for bidirectional design, an approach that lets the assistant speak, hear, and listen at once rather than waiting politely for each turn. Internal code reportedly frames it as the next generation of voice and a major leap in intelligence.
Early testers say the model has already begun reaching a subset of users across web and mobile, which signals a release as soon as this week, though the final name may still change.
Also Read: Is The Anthropic Perp Sell-Off A Warning For Pre-IPO Crypto Bets?
Bidi 1 Handles Interruptions And Memory
The model offers small acknowledgments, like a quiet "okay," when a user pauses or slows down, and it manages that without cutting the speaker off. It can switch tasks on the fly, reversing a count the moment a user interrupts. Reports describe selectable intelligence tiers labeled High, Medium, and Instant, mirroring the choices already offered on the text side, where users pick faster or more careful answers.
Memory may prove the bigger shift, since Bidi 1 holds the thread of a long conversation instead of dropping earlier audio context, the weak point that has long dogged ChatGPT's current voice stack. One sighting even pointed to real-time translation, a feature that could unlock fresh use cases once the model reaches the developer interface and powers outside apps.
OpenAI Voice Push Gathers Pace
The upgrade reads as a bid to close the gap between OpenAI's strong text models and an older voice layer that has lagged behind for months. That layer leaned on GPT-4o, a model never built from the ground up for two-way audio. The company is betting that speech, not typing, becomes the main route into AI for most people.
OpenAI has refined ChatGPT's voice features steadily over the past year, and the model has reportedly been under development since early 2026, the product of months of work rather than a rushed release. The leak also lands as the company maps a wider ChatGPT overhaul around its Codex coding tool and agentic features, though none of it is official yet.
Read Next: Mane City Mobile Lands On iOS And Android In 100+ Countries