OpenAI appears to be building a real-time voice mode for Codex that extends well beyond code. Recent additions to the ChatGPT bundle include strings and supporting functionality that describe a general-purpose agentic assistant you can communicate with via phone. The voice layer maintains the conversation while worker agents perform tasks in the background. The system prompts behind it reference checking Slack, pulling up documents and calendars, controlling Spotify, browsing, shopping, and food ordering, including scanning Uber Eats for dinner. This represents a personal assistant framework rather than a developer one, and it operates on top of Codex Remote Control. Thus, voice would become a means to dispatch and manage several tasks simultaneously on a laptop, with results read back to the user.
— M1 (@M1Astra) July 23, 2026
Several aspects remain uncertain. It is unclear which model will handle the conversations, with the bidirectional voice model spotted earlier this summer being the obvious candidate, though nothing has been confirmed yet. It is also uncertain whether Codex Real-Time will have its own entry in the voice model selector. A real-time voice section was observed inside the Codex app before the work was restructured more broadly, and it has not appeared on the ChatGPT desktop app, possibly residing in a bundle that has not been released.
Unbelievable excited for what’s coming together. Tomorrow is feeling codexy
— Tibo (@thsottiaux) July 23, 2026
The framework aligns with the company's direction. OpenAI recently integrated the Codex app into the ChatGPT desktop app, positioning Codex alongside Chat and Work. Codex Remote reached general availability across paid plans in June. A spoken layer that covers both coding threads and everyday tasks suits a company that is consolidating its interfaces into one assistant, potentially attracting users who never open a terminal.
The timing is uncertain, and this may be early groundwork. Codex staff have hinted at something for tomorrow, with discussions pointing to parallel worker agents and faster Cerebras-served inference for GPT-5.6 Sol, rather than a new silicon deal, since that partnership has been ongoing since January and has already produced Codex-Spark. Whether the voice feature will be included in the same release is the aspect worth monitoring.
