Developer @DevAdventur3s recently dug up 1,536 lines of unactivated Rust code from OpenAI's Codex repository, revealing an upcoming real-time voice mode in internal testing. The big change: a complete separation of interaction and execution, allowing a front-end voice call to run in parallel with back-end code writing.
https://twitter.com/DevAdventur3s/status/2055765342484590774
According to leaked UI and source comments, after a user issues a complex command like refactoring via voice, the front-end immediately fires up a voice model codenamed gpt-realtime-1.5. It uses WebRTC to have a real-time conversation with the user and verbally report progress. Meanwhile, the heavy lifting — pulling files, modifying code, running tests — is silently handled by a larger model in the background.
AI programming's interaction model is moving from turn-based text chat toward something that feels like a real-time call with a pair-programming coworker. The underlying logic and supporting UI have already been merged into the main branch. All that's left is for OpenAI to flip the switch on the server side.