OpenAI looks set to bring real-time voice to Codex, and the way it is arriving says as much about branding as about capability. A voice control that once did nothing now summons the animated pet the coding app gained earlier this year, a sign that the session handling has come online ahead of a public switch. The timing lines up with separate signals that a next-generation voice model, tentatively tagged GPT-Bidi-1, could reach ChatGPT's voice mode next week, built around a bidirectional design that listens and speaks at once rather than taking strict turns.

0:00

/0:33

Inside the app, a real-time voice section now includes controls for developers who want to keep the spoken channel open while code runs. One assigns a hotkey and a wake word, with sessions starting on the phrase "Hey Chat." A single-tone option pins new sessions to one durable orchestrator thread rather than spawning fresh ones, letting a developer return to the same continuous context.

A third setting governs the avatar overlay, an Orb or the Pet, though the Orb has yet to render since the Pet keeps appearing. A Library entry has also surfaced in the sidebar, not yet openable but closely modeled on the one ChatGPT already uses.

That borrowing is the real story. The orb, the library, and a wake word that says "Chat" rather than "Codex" suggest OpenAI treats its coding agent and consumer assistant as one converging surface, not two products. For a company betting speech becomes the main route to its models, a single surface is the point. Whether the Codex side flips the same week as the ChatGPT model or trails behind, which is plausible, the direction is set.

Anthropic is working on the same ground. Multilingual support has begun appearing in Claude's voice mode without a formal announcement, alongside a push-to-talk option, and it reads as a first layer rather than the full plan, with the underlying model and system likely due for an upgrade, too. The quiet rollout fits a habit of holding the louder reveal until a rival forces the moment.