OpenAI has added ChatGPT Voice to its desktop application, letting users give complex spoken instructions that trigger multiple steps and receive spoken prompts when the assistant needs more information. The feature went live on July 23, 2026, with the official announcement following on July 24 (TechCrunch).

The desktop integration runs on OpenAI's ChatGPT-Live voice model family, which launched earlier in July 2026. Voice in the desktop app works with both ChatGPT Work and Codex (OpenAI's coding assistant), and can use computer-use capabilities to look up websites and applications. On macOS, a feature called Appshots gives ChatGPT Voice access to what is on the user's screen, including alt-text — the text descriptions embedded in images and interface elements (TechCrunch).

The feature set goes beyond conversational voice. OpenAI published a demo video showing a developer using ChatGPT Voice to create a new thread, submit a pull request (a proposed code change), and find the root cause of a bug, all in a single spoken command. The desktop version responds out loud when it needs input, allowing a back-and-forth workflow that does not require the user to return to the keyboard between steps (TechCrunch).

For Codex users, Voice is also accessible remotely from the iOS app. A developer can start a voice-driven Codex session from a phone without sitting at the machine where the work is running (TechCrunch).

This is not the first time OpenAI has offered voice in the macOS desktop app. A previous Voice experience was retired on January 15, 2026 (OpenAI Help). The smartphone version of ChatGPT Voice that launched with the ChatGPT-Live family delivered smoother conversations and better interruption handling, but was not built to take actions on the device. The desktop version closes that gap by pairing voice with the computer-use tools already available in the desktop client.

The release lands against an active competitive backdrop. On or around July 23, 2026, Anthropic updated Claude's voice mode to route through its Opus, Sonnet, and Haiku models for completing tasks across Gmail, Calendar, Slack, Notion, and Canva (TechCrunch). Both companies are converging on the same idea: voice as a control surface for agentic workflows — automated sequences where the assistant chains together multiple tool actions on the user's behalf — rather than just a conversational interface.

The distinction matters. The smartphone voice experience optimized for natural turn-taking and quick recovery when a user interrupts. The desktop version optimizes for task execution, chaining multiple tool invocations behind a single spoken instruction and using screen context to inform its actions. The Appshots capability on macOS is particularly notable: by reading on-screen content including alt-text, the assistant can ground its responses in what the user is actually looking at, reducing the need to verbally describe the working context.

A demo showing a clean thread-creation, pull-request, and bug-diagnosis chain in one command is compelling, but agentic workflows have a well-documented tendency to fail in unanticipated ways when the environment deviates from the expected path. Users adopting this in production will want to understand the failure modes before relying on it for time-sensitive work.

The remote access angle for Codex via iOS is worth a closer look. It extends the desktop voice experience beyond the desk, letting a developer interact with a running Codex session while away from the machine. For teams already using Codex for background code tasks, this could shift how and when they engage with those sessions, turning a periodic check-in into a continuous voice channel.

The competitive dynamic with Anthropic adds pressure on both execution quality and ecosystem breadth. Claude's voice mode currently targets productivity apps like Gmail and Slack. OpenAI's desktop approach targets developer workflows through Codex and screen-level computer use. The two strategies are not yet directly overlapping, but they are closing in on the same user expectation: speak a task, and the assistant carries it out across whatever tools are necessary.

The previous desktop Voice experience was absent for roughly six months between its January retirement and this release. That gap suggests OpenAI treated the earlier version as insufficient and chose to rebuild on the ChatGPT-Live foundation rather than iterate incrementally. The result pairs a newer speech model with the computer-use infrastructure the desktop app already had, which is a meaningfully different proposition from the earlier Voice feature.