📰 Key Highlights

OpenAI announced on Thursday that it’s adding voice support to the ChatGPT desktop app, letting users control AI agents through conversation and execute tasks on their computer. This new feature uses OpenAI’s latest voice model series “ChatGPT-Live” released earlier this month. OpenAI says ChatGPT Voice works alongside ChatGPT Work and Codex, and can invoke computer-use skills to query websites and apps. On macOS, through the Appshots feature, users can also have apps read screen content, including alt-text. When voice mode first launched on mobile, it focused on smoother conversational flow and better interruption handling, but it didn’t actually execute operations on the phone at that time. This desktop update is more complete — users can dictate complex multi-step instructions, and ChatGPT can respond in real time when it needs user input. In a demo video released by OpenAI, a developer used a single command to have ChatGPT create a new thread, submit a pull request, and find the root cause of a bug. The company also noted that users can use ChatGPT Voice with Codex via remote connection through the iOS app. Separately, Anthropic simultaneously updated Claude’s voice mode, which can invoke Opus, Sonnet, and Haiku models to complete tasks in apps like Gmail, Calendar, Slack, Notion, and Canva.

💬 JudyAI Lab’s Take

OpenAI adding voice-controlled AI agent capability to the ChatGPT desktop is worth noting — it’s no longer just about “voice responses,” but actually completing multi-step tasks hands-on on your computer. At the same time, Anthropic’s Claude voice mode is following suit, hooking into common tools like Gmail, Calendar, and Slack.

This reflects the trend of voice interfaces shifting from an “input method” to an “operation interface” — in the past, voice assistants mainly solved the hassle of typing, but now they’re letting AI agents directly take over execution, including writing code, debugging, and cross-app queries. When voice, screen understanding (like Appshots reading the screen), and tool calling combine, what AI builders need to think about is no longer just “are the answers accurate,” but “can the agent safely and controllably complete multi-step operations” — including interruption handling and user confirmation mechanisms. The experience design of these voice-driven agents will become one of the key product differentiators in the next phase.

If you’re also building AI tools, think about this: which workflow in your product is suitable for being converted into a voice-driven, multi-step agent?

📅 Source Info

🔗 Further Reading