OpenAI launches GPT-Live voice models with full-duplex conversation and simultaneous web search

OpenAI released new voice models for ChatGPT on July 9, 2025, according to ZDNET testing. The models use full-duplex architecture that allows ChatGPT to speak and listen at the same time—a departure from previous voice modes that could only handle one task at a time.

Two models for different user tiers

GPT-Live-1 is the default voice model for ChatGPT Go, Plus, and Pro subscribers. Free users get GPT-Live-1 mini. According to OpenAI, GPT-Live-1 offers higher quality while the mini variant uses fewer resources. Pricing for the voice models has not been separately disclosed—they're included in existing ChatGPT subscription tiers.

Technical capabilities

The full-duplex architecture allows the AI to provide conversational feedback with interjections like "Yeah" or "Mhmm" while users speak. Users can interrupt the AI mid-sentence, and it adjusts in real-time.

A key feature: ChatGPT can conduct web searches during voice conversations. The system hands off search queries to a separate model, allowing the voice conversation to continue uninterrupted. This parallel processing enables users to ask research questions without breaking conversational flow.

The models include three reasoning levels selectable through an "Intelligence" setting: Instant (default, for quick replies), Medium (deeper analysis with longer response times), and High (complex problem-solving with longest latency).

Platform availability

GPT-Live is available on:

• ChatGPT website
• Windows app
• iOS app
• Android app

The ChatGPT Mac app no longer supports Voice mode. Mac users must use the web interface.

Testing results

ZDNET testing confirmed the models handle interruptions smoothly across multiple use cases: real-time story modifications, translation between English and French, and technical troubleshooting with mid-conversation web searches.

Two ZDNET editors reported their conversations were interrupted by background audio and nearby speakers. A third editor could not replicate this issue with loud TV dialogue and background voices. OpenAI has not commented on background noise handling.

What this means

This release represents a significant technical step in making voice AI interactions more natural. The ability to maintain conversation while executing background tasks—particularly web search—addresses a major UX friction point in current voice assistants. However, the bifurcation into full and mini models based on subscription tier follows OpenAI's standard approach of feature gating. The background noise issue reported by some testers suggests the models may need refinement in audio filtering, though inconsistent reproduction makes this unclear.