On July 8, 2026, OpenAI officially unveiled GPT-Live, a new generation of voice model for ChatGPT's voice mode, and began rolling it out. Its defining trait is a full-duplex architecture that lets the AI listen while it speaks, handling user interruptions and backchannel cues in real time.
July 8, 2026 · OpenAI
GPT-Live: The Voice AI That Listens While It Speaks
OpenAI unveils a full-duplex voice model for ChatGPT — handling interruptions, overlap and backchannels in real time, in two variants rolling out globally.
2
Variants at launch — GPT-Live-1 and GPT-Live-1 mini
Full-duplex
Listens and speaks simultaneously, not turn-by-turn
Global
Rolling out to ChatGPT users worldwide
Turn-based vs. Full-duplex
How much of the conversation can flow at once (relative capability)
Turn-based
Prior voice mode & Realtime series — no mid-speech interruption, pauses & latency
Full-duplex
GPT-Live — barge-in, overlap, backchannels & silent waiting, all live
What's new in GPT-Live
Real-time barge-in
Interrupt mid-speech; "mhmm" / "yeah" backchannels & natural pause detection
Live translation & emotion
Expanded emotional tone, stronger context retention, mid-talk task switching
Frontier delegation
Hands complex search / reasoning to a model like GPT-5.5, weaving results back in
The lineup
GPT-Live-1
Paid tiers — Go / Plus / ProAPI coming soon
GPT-Live-1 mini
Free usersAPI coming soon
Reaction — largely positive
Called strikingly natural and human-like — a clear step up in quality and latency. Some see voice becoming a primary interface, with translation and agentic use as concrete cases.
The open question
No model-specific benchmarks — such as Full Duplex Bench scores — have been published, leaving quantitative validation of performance for later.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article…