OpenAI launches GPT-Live: simultaneous listening and speaking, creating a human-like conversation experience

OpenAI launched full-duplex voice models GPT-Live-1 and GPT-Live-1 mini on Wednesday, replacing the old Advanced Voice Mode. Complex tasks will be delegated to GPT-5.5, released this April.

(Previous: Google Translate upgrade: Gemini 3.5 eliminates awkward pauses in real-time voice interpretation) (Background: Build an AI customer service agent in 2 minutes! xAI launches no-code Voice Agent Builder, voice benchmark beats OpenAI)

Ending two years of "turn-based conversation," OpenAI taught ChatGPT one thing on Wednesday: listen and talk simultaneously. The new voice model GPT-Live abandons the old logic of judging turns by silence, adopting a full-duplex architecture.

GPT-Live-1 and GPT-Live-1 mini are now available globally on iOS, Android, and ChatGPT.com. GPT-Live-1 becomes the default voice model for Go, Plus, and Pro paid users, while GPT-Live-1 mini serves free users. The API will require developer waitlist access. This is the third overhaul of ChatGPT's voice technology in two years, and OpenAI's clearest statement yet: voice is no longer an add-on to a text chatbot, but the primary interface for human-AI conversation.

Why did the old voice always interrupt?

OpenAI wrote in a research blog that GPT-Live continuously processes input while generating output. The model makes multiple interaction decisions per second, judging whether to speak, keep listening, pause, interrupt, or call tools. A researcher on X described the old experience as "turn-based," where only one person could speak at a time, and you had to wait for the other to finish.

This was the Achilles' heel of the old Advanced Voice Mode. OpenAI admitted that because it relied on silence to judge turns, brief pauses or background noise could be mistaken for the user finishing, causing the model to interrupt at unnatural moments.

Who speaks, who thinks?

Another key design of GPT-Live is separating the "voice interaction layer" from the "reasoning layer." Simple questions are answered directly by GPT-Live. For tasks requiring web searches, deep reasoning, or multi-step processes, it delegates the work to a frontier model running in the background—currently GPT-5.5, released this April. During computation, it continues conversing with the user, without the seconds-long blank periods of the old pipeline.

This modular design has a benefit: upgrading intelligence no longer requires retraining the voice model itself. For businesses and developers who want to connect voice agents to databases, search, and multi-step tasks, this truly lowers the barrier.

The new version also adds two small features: visual cards for weather, stock prices, sports scores, maps, etc., appear in real-time during conversations. Users can also switch between three reasoning levels: Instant (immediate), Medium (moderate), and High (complex), deciding for themselves whether they want faster voice response or more thorough answers. In other words, speed and accuracy are no longer a trade-off; users can draw that line themselves.

OpenAI also revamped its voice lineup, with nine voices emphasizing "designed for conversation, not voice imitation," and added protections against mimicking real human voices. This is partly a response to an old controversy from two years ago. In May 2024, when GPT-4o was launched, the voice "Sky" was widely perceived as resembling Scarlett Johansson's voice from the film Her. Johansson confirmed she had declined Sam Altman's voiceover invitation and said she was shocked and angry after the product launched. OpenAI later removed that voice and apologized.