OpenAI is secretly developing a real-time voice mode for Codex

According to Dynamic Beating monitoring, developer @DevAdventur3s recently uncovered 1536 lines of inactive Rust code from the OpenAI Codex codebase, exposing Codex's real-time voice mode currently in internal testing. The biggest change in this update is a complete separation of interaction and execution, achieving a dual-path parallelism of live collaboration and background coding.

From the leaked interface and source code comments, it appears that after users issue complex instructions such as refactoring via voice, the frontend immediately invokes a model codenamed gpt-realtime-1.5 for voice interaction, engaging in a real-time conversation with the user via WebRTC to verbally report progress. Meanwhile, the heavy lifting of file retrieval, code modification, and test execution is silently handled by another backend model with a larger parameter size.

The interactive experience of AI programming is transitioning from turn-based text Q&A to real-time conversations akin to pair programming with a colleague. Currently, this underlying logic and corresponding UI have been merged into the main codebase, awaiting only the activation by OpenAI's server-side permission to go live.