OpenAI officially launched the full-duplex voice models GPT-Live-1 and GPT-Live-1 Mini on July 9, now available on iOS, Android, and ChatGPT.com worldwide, replacing the old Advanced Voice Mode.
GPT-Live-1 and GPT-Live-1 Mini launch specifications: paid users, free users, and API queue arrangements
According to OpenAI's announcement, the scope of the two GPT-Live models is as follows: GPT-Live-1 is the default voice model for Go, Plus, and Pro paid plan users; GPT-Live-1 Mini serves free users; for API access, developers are currently on a waitlist and not yet available.
Both models are now live globally on iOS, Android, and ChatGPT.com, marking the third revision of ChatGPT's voice technology within two years and representing OpenAI's most explicit positioning statement: voice will become the main interface for human-AI interaction, rather than an auxiliary feature of text chatbots.
Principles of full-duplex architecture: issues with silent turns in the old Advanced Voice Mode and GPT-Live's solution
According to OpenAI's research blog, the core issue with the old Advanced Voice Mode was relying on silent detection to determine turns—short pauses or background noise could be mistaken for the user finishing speaking, causing the model to interrupt unnaturally; this "carousel-style" turn-taking made conversations feel unnatural, with only one person speaking at a time.
GPT-Live adopts a full-duplex architecture, allowing the model to process input continuously while generating output, making multiple interaction decisions per second—such as when to speak, continue listening, pause, interrupt, or call tools—closer to real human conversation.
Separation of voice layer and reasoning layer: GPT-5.5 delegation, three-level reasoning, and new features
The key design of GPT-Live is the modular separation of the "voice interaction layer" and the "reasoning layer": simple questions are handled directly by GPT-Live; complex tasks (like web searches, deep reasoning, or multi-step tasks) are delegated to GPT-5.5 running in the background, maintaining ongoing dialogue without the several-second pause typical of previous pipelines.
This modular design allows upgrades to reasoning intelligence without retraining the voice model. Companies and developers can connect voice agents to databases, search functions, and multi-step tasks. New features introduced simultaneously include:
-
Three-level reasoning: Instant (real-time), Medium, and High (complex), allowing users to choose their preferred balance of speed and depth.
-
Real-time visual cards: Automatically display information cards such as weather, stock prices, sports scores, and maps during conversations.
-
Nine voice options: Redesigned voices emphasizing "designed for dialogue, not voice imitation," with mechanisms to prevent voice mimicry.
Frequently Asked Questions
What are the differences between GPT-Live-1 and GPT-Live-1 Mini plans?
According to OpenAI, GPT-Live-1 is the default voice model for Go, Plus, and Pro paid plan users; GPT-Live-1 Mini serves free users; API access is on a waitlist for developers. Both models are available immediately on iOS, Android, and ChatGPT.com worldwide.
How does GPT-Live handle complex tasks, and how does it differ from the old Advanced Voice Mode?
OpenAI's design explanation states that GPT-Live uses a modular architecture separating voice interaction and reasoning layers; complex tasks are delegated to GPT-5.5 in the background, with continuous dialogue during processing, avoiding the several-second pauses of the old version. The old Advanced Voice Mode relied on silent detection, which could misjudge turn-taking, whereas GPT-Live's full-duplex approach allows simultaneous input processing and output generation.
Why does the new GPT-Live voice design emphasize "preventing voice mimicry"?
According to reports, this design is related to the "Sky" voice incident in May 2024 when GPT-4o was announced: the voice was believed to resemble actress Scarlett Johansson, who had previously refused a voice-over invitation from Sam Altman and later described being shocked and angry. OpenAI withdrew that voice and apologized. The nine new voices emphasize "designed for dialogue, not voice imitation," with added technical protections against voice mimicry.
