OpenAI Replaces ChatGPT's Voice With a Model That Listens and Talks at Once

OpenAI has swapped out the voice engine behind ChatGPT. As of July 8, 2026, Advanced Voice Mode is being replaced by GPT-Live-1, the company's first model that listens and speaks at the same time instead of waiting for the other person to finish talking. The change rolls out to paid and free users worldwide, including in Poland.

The key change is architectural. The previous voice mode worked in turns: the user spoke, the model waited, then it replied. GPT-Live-1 runs in full duplex, meaning it processes incoming audio and generates a response at the same time, the way a person does in a natural conversation.

How full duplex works

In practice, this means the model can drop in short acknowledgments like "mhm" or "sure," jump into a sentence mid-thought, or deliberately stay quiet when it senses the other person is still gathering their thoughts. OpenAI describes this as an attempt to recreate the rhythm of a real conversation, where nobody politely waits for a signal that the other person is done talking.

For tasks that require deeper reasoning, web search, or more complex work, GPT-Live-1 doesn't try to handle everything on its own. Behind the scenes it delegates such questions to GPT-5.5, while the voice layer's job is simply to keep the conversation flowing smoothly while it waits for an answer.

The benchmark numbers

OpenAI published comparative results showing the biggest gains in reasoning and search tasks. On GPQA, which tests graduate-level scientific reasoning, the model in high-reasoning mode hit 84.2 percent accuracy versus 45.3 percent for Advanced Voice Mode.

An even bigger gap shows up in BrowseComp, a benchmark that tests agentic web search ability. GPT-Live-1 scored 75.2 percent there, the mini version 31.6 percent, and the previous model just 0.7 percent. On the tau3 Voice Telecom test, which simulates multi-turn conversations with telecom customer support, the high-reasoning variant hit around 65 percent accuracy with an average call time of 385 seconds, compared to about 30 percent for the previous version.

Enthusiasm that grates

The launch has been met with a wave of criticism on social media. Users testing GPT-Live-1 note that the model is a little too eager to chime in with short acknowledgments and jump in before the other person finishes a thought. Short reactions like "mhm" or "got it," meant to sound natural, have become, for some users, a new form of interruption.

The outlet The Decoder points to a key difference between text and voice assistants: sound intrudes on a person's attention in a way that text on a screen never does. A model that would come across as helpful in text can feel intrusive in a voice conversation.

What it means for business

For businesses, including in Poland, the customer-service benchmark results and the scale of rollout matter most: 150 million weekly voice ChatGPT users give OpenAI a base to quickly test whether full duplex actually shortens service calls or just complicates them. Companies planning voicebots built on OpenAI's API get a model with genuinely higher accuracy in multi-turn support scenarios, but also a reputational risk if the assistant keeps interrupting customers.

For now, GPT-Live-1 doesn't support voice combined with video or screen sharing in ChatGPT, features that were available in the older voice modes. OpenAI says it's working on adding them, and users who need video can, for now, switch back to the older Standard and Advanced Voice Mode.

The GPT-Live-1 launch is part of a broader race to make voice the interface for AI agents, where OpenAI competes with, among others, Google's Gemini Live. How users respond to the new model's over-eagerness in the coming weeks may determine whether full duplex becomes the standard, or whether it gets dialed back to a more restrained version.

Sources: Introducing GPT-Live (openai.com), ChatGPT can now listen and talk at the same time (the-decoder.com), OpenAI Releases GPT-Live and GPT-Live-1 mini (marktechpost.com), OpenAI Launches GPT-Live Voice Model (finance.biggo.com)