OpenAI added GPT-Live-powered ChatGPT Voice to version 26.715 of its macOS and Windows desktop app on July 23, 2026, letting eligible users start, check, and steer work in Chat, Work, and Codex by speaking. Plus, Pro, Business, Edu, and Enterprise plans are included, while macOS users can optionally share the frontmost window through an appshot.
OpenAI added GPT-Live-powered ChatGPT Voice to version 26.715 of its desktop app for macOS and Windows on July 23. The feature lets users speak with ChatGPT while working in Chat, Work, and Codex, including starting tasks, checking their status, and issuing follow-up instructions by voice.
OpenAI's Voice documentation says a user must begin a new chat or task in voice mode to use the full live interaction. A chat that starts in another mode offers voice dictation instead. Once a voice session is active, users can interrupt a response, ask follow-up questions, or redirect work in progress.
Voice across agentic workflows
The three desktop experiences have different scopes:
- • Chat handles conversational assistance, search, brainstorming, and everyday questions. - • Work is designed for longer tasks such as research and creating documents, spreadsheets, presentations, reports, or sites. - • Codex focuses on software development, including writing or debugging code, running tests and commands, reviewing changes, and working with repositories.
OpenAI says Voice can start separate threads for longer tasks, inspect existing threads, and bring progress, blockers, and results back into the voice conversation. The official documentation gives examples such as asking Codex to run tests and investigate failures, then checking active tasks for blockers.
The July 23 changelog describes the feature as a way to start, check, or steer work in other threads. Independent reports from 9to5Mac and Neowin also confirmed the macOS and Windows rollout and its connection to Work and Codex.
Screen context on macOS
On macOS, users can enable Screen context in Voice settings and ask ChatGPT to inspect the frontmost window. OpenAI calls the resulting capture an appshot. Its documentation says an appshot may contain both the window image and accessible text, including text outside the visible scroll area. Organizations can disable the capability.
That permission model matters for enterprise deployments because the capture may expose more than the pixels currently visible. OpenAI advises users to avoid sharing windows that contain sensitive information, and macOS may request Screen & System Audio Recording and Accessibility permissions.
Availability and practical limits
The desktop feature is available for Plus, Pro, Business, Edu, and Enterprise plans, subject to rollout status and workspace settings. Enterprise and Edu access begins with a two-week early-access period before default availability. Users can also access the experience through Remote on iOS after pairing a phone with a desktop host.
For teams evaluating agent interfaces, the important change is not dictation alone but a voice layer for coordinating task-oriented Work and code-oriented Codex workflows. OpenAI says Voice follows the tools and permissions of the selected experience, so task state, usage limits, approvals, and screen-context controls remain operational considerations rather than disappearing behind the spoken interface.
Key Points #
- 1ChatGPT desktop app version 26.715 adds GPT-Live Voice across Chat, Work, and Codex on macOS and Windows.
- 2Voice can start, check, and steer work in other threads while retaining the tools and permissions of the selected experience.
- 3On macOS, optional Screen context can share an appshot of the frontmost window, including accessible text outside the visible scroll area.
Scoring Rationale #
The update connects GPT-Live voice interaction to task-oriented Work and developer-oriented Codex workflows in the desktop app. It is operationally relevant for teams evaluating agent coordination, desktop permissions and hands-free software workflows, while remaining an extension of existing ChatGPT and Codex capabilities.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.