ChatGPT Voice offers a natural conversation experience with full duplex audio

Summary

OpenAI released GPT-Live in July 2026 for mobile and desktop platforms. The system offers full-duplex voice interaction. Users can interrupt mid-sentence, change topics, and seamlessly blend voice with text. Agent capabilities are also integrated via Work (documents, tables) and Codex (code, local projects) components.

Why it matters

This marks the shift from sequential chatbot Q&A to a continuous, context-aware assistant that performs tasks like browsing, file editing, and multi-step operations on your behalf while you are busy with other work. This also increases competitive pressure on Apple's Siri. It is reported that Siri will integrate third-party models (ChatGPT, Claude, Gemini) in iOS 27. However, the company chose not to announce this at WWDC. This preference highlights how uncertain the voice assistant market still is.

Highlights

• Full-duplex voice allows both parties to speak simultaneously. The system detects when you are done, correcting yourself, or when it should remain silent.
• Consequential actions such as buying/selling, emailing, changing permissions, and transferring money require explicit user confirmation at every step by design.
• A ChatGPT Plus or higher subscription is required. The free tier offers a limited voice mode without agent capabilities.

Critical take

The article relies largely on a single New Atlas report and the author's own enthusiasm. However, the underlying model still 'confabulates.' That is, it may silently misprice a tire or file a document in the wrong folder while you are not looking. While the step-by-step confirmation mechanism is logical, it does not cover every error scenario. Privacy areas such as screen access, calendar, and linked accounts are acknowledged. However, OpenAI has not published a detailed threat model analysis for such an intrusive product. This is a notable omission.