Agentic coding goes hands-free as OpenAI brings GPT-Live's full duplex voice control to Codex and ChatGPT on…

Two weeks after debuting GPT-Live, its more naturalistic audio AI model, OpenAI is wiring that same full-duplex voice stack into developer workflows. The company announced that GPT-Live now powers the ChatGPT desktop application and extends full-duplex voice control to Codex, so agentic coding can run hands-free: the model listens and speaks at the same time rather than waiting for a turn to end.

Full duplex is the core product mechanic. Unlike half-duplex voice UIs that block listening while speaking, GPT-Live is built so the system can hear the user while it is still talking. In the ChatGPT desktop app that means continuous, conversational control of the session; in Codex it means voice can drive agentic coding actions without a push-to-talk cadence or a forced pause between instruction and response.

For engineers and builders, the shift is operational, not cosmetic. Hands-free control on the desktop lets you steer an agent while reading code, watching a build, or switching windows. When voice no longer requires stopping the model to interrupt or wait for a clean turn boundary, you can correct course mid-task, add constraints on the fly, and keep the agent in the loop without dropping to the keyboard for every micro-instruction.

Competitive pressure around voice-first and agentic coding tools has been rising as vendors race to make assistants feel less like dictation and more like a co-worker in the same room. OpenAI’s move folds a consumer-facing, naturalistic audio model into the same desktop surfaces developers already use for ChatGPT and Codex, tightening the gap between general chat voice demos and tools that actually touch repositories, diffs, and multi-step coding agents.

What to watch next is how far voice reaches into the agent loop: which Codex actions accept interruptible speech, how the desktop app handles noisy environments and overlapping talk, and whether full-duplex control stays a desktop privilege or shows up in the same form across other OpenAI coding surfaces. The concrete test is whether builders use it for real multi-step work, not just demos—continuous steering of an agent without abandoning the keyboard when precision still matters.