OpenAI brings full-duplex gpt-live voice to Codex and ChatGPT desktop

Transformative voice interaction has officially arrived at the desktop operating system level. OpenAI has rolled out a major desktop update to its macOS and Windows applications, integrating its full-duplex "GPT-Live" voice architecture across ChatGPT Work, desktop environments, and the Codex software development interface.

Editorial credit: Koshiro K / Shutterstock

The technical upgrade represents a fundamental shift away from legacy, turn-based speech interfaces. Rather than treating conversations like a walkie-talkie, where users must wait for an AI model to finish speaking before replying, the GPT-Live engine processes incoming audio and generates spoken output continuously and simultaneously. The resulting full-duplex interaction allows users to interrupt the model mid-sentence, pause to think without triggering accidental responses, or issue real-time verbal redirects while the system actively works.

Directing Autonomous Agents by Voice

By bringing full-duplex voice control directly into the ChatGPT desktop ecosystem, OpenAI aims to establish speech as a primary command interface for multi-threaded work. Users can now direct, check, and adjust background AI operations across active projects using natural conversational steering:

• Codex Integration: Software developers can review codebases hands-free, direct background test suites, or instruct Codex agents to debug software flaws while discussing high-level software architecture.

• Agentic Coordination: Users can initiate separate background threads for complex research tasks and receive verbal updates, progress reports, or approval requests without switching windows or typing prompts.

• Desktop Appshots: On macOS, the feature introduces visual context awareness via "Appshots," enabling the model to reference active, frontmost window content, including accessible text outside the visible scroll area, when users ask for input.

Background Delegation via GPT-5.5

To handle complex logical tasks without stalling real-time vocal output, the architecture separates natural conversation from heavy computational reasoning. When a user asks a complex question, initiates a heavy coding sequence, or requests a web search during a voice session, GPT-Live delegates the heavy lifting to frontier reasoning models (such as GPT-5.5) in the background. The voice model keeps the conversation flowing smoothly before seamlessly weaving the generated results back into the spoken chat once complete.

The desktop voice capabilities are rolling out globally to ChatGPT Plus, Pro, Business, Edu, and Enterprise plan subscribers. By converting natural speech into an active control layer for autonomous agents, OpenAI is taking a significant step toward embedding hands-free, conversational orchestration directly into everyday professional workflows.