GPT-Live, the newest generation of voice models created to make interaction with ChatGPT more human-like, has already rolled out from OpenAI. These models are capable of delegating a question to a frontier text model to perform a web search or resolve a difficult situation right within a voice conversation.
The rollout began on July 8 and is expanding to reach all ChatGPT Voice users worldwide via iOS, Android, and Web.
What OpenAI Launched
There are two different versions of the model available at the moment – GPT-Live-1 and GPT-Live-1 mini. While the former has become the default model powering ChatGPT Voice for Go, Plus, and Pro users, the latter is now the default for Free users. Both of the models are called full-duplex, which means that they allow receiving audio and generating speech at the same time instead of using turn-based communication.
As it turns out, the difference between full-duplex and turn-based communication is quite substantial. Previously developed voice systems only responded after the person finished speaking, since they couldn’t process speech and generate a response at the same time. Full-duplex technology makes it possible to give some feedback to the user “got it”, etc.), to wait for some time while the person collects his thoughts, and to respond to the interruption immediately without delay.
Before this release, there had already been two phases of ChatGPT Voice development. The first version of the model was based on chaining three different models – speech-to-text, language model, and text-to-speech model. This model allowed for real-time conversation, although it felt unnatural since all the information passed through three models in a row. Another phase of development was Advanced Voice Mode which was able to process and generate audio inside the model itself, which significantly reduced the delay and made the conversation smoother. However, this model still used turn-based communication and detected the end of a turn via silence. Because of this, it could make mistakes when there was background noise or a short pause from the person. So the new GPT-Live model aims to solve both of these problems by using a continuous architecture, which allows the model to make decisions regarding listening, speaking, pausing, and using a tool many times per second.
Search and Deeper Reasoning Move Into the Conversation
Another important feature of GPT-Live is its ability to handle tasks that require more effort than a simple reply. Instead of handling it alone, GPT-Live can delegate the task to a separate frontier model, GPT-5.5, at launch, while the conversation continues. After getting the answer, the model will insert it into the voice conversation.
This represents a significant architectural improvement over previous models, since ChatGPT Voice is now split into two parts that work separately. GPT-Live is responsible for making interaction natural, and GPT-5.5 is working on searching and deep reasoning itself. As a result, GPT-Live can automatically keep up with the latest frontier models without rebuilding the voice experience. Users can choose how much reasoning effort to apply to a given question. Instant mode, built for quick answers, runs on GPT-5.5 Instant, while Medium and High modes use GPT-5.5 Thinking.
Additionally, ChatGPT Voice now displays visual cards on the topics like weather, stocks, sports scores, and maps even while the person is still talking. So the spoken answer will get a visual supplement instead of switching to the text mode. OpenAI says Voice continues to support features like search, memory, image upload, and file upload, just as it did before.
How OpenAI Says It Measured the Improvement
OpenAI created special human evaluations to assess how pleasant and natural the voice conversation feels, instead of relying on traditional benchmark scores. Compared with Advanced Voice Mode in conversations of five to ten minutes, OpenAI claims that GPT-Live-1 and GPT-Live-1 mini are highly preferred on measures like overall preference, turn-taking, interruptions, conversational flow, and naturalness of interaction.
At the same time, OpenAI claims some improvements regarding traditional benchmarks. For example, on GPQA which evaluates expert-level scientific reasoning in biology, chemistry, and physics, GPT-Live-1 significantly outperforms Advanced Voice Mode. There are strong improvements on BrowseComp benchmark which evaluates agential web search and finding of the rare information. GPT-Live-1 outperforms Advanced Voice Mode on an internal benchmark which evaluates voice agents on multi-turn telecom support scenarios.
To provide the scale of these results, OpenAI says that more than 150 million people talk to ChatGPT each week using such features like Voice and Dictation.
Safety Measures Built Specifically for Voice
Since conversations via voice involve real-time engagement and cover highly sensitive topics, OpenAI explained that GPT-Live was designed with specific safety training and additional safety safeguards compared to those applied to the company’s text models. The company added safety testing and audio-specific safety tests to its arsenal, as well as synthetic audio evaluations on self-harm, psychosis and mania, emotional reliance on AI, violence, and sexually explicit content based on the lessons from Advanced Voice Mode. According to the company, GPT-Live outperformed Advanced Voice Mode in almost all of the areas listed above.
In addition, the company mentioned that certain safeguards could kick in while the model is still mid-conversation. In case of any potentially dangerous output, the system can redirect the model to give a safe answer, surface safety messaging or relevant resources, or stop a voice conversation in higher-risk cases. Regarding self-harm, OpenAI has implemented existing support flows in ChatGPT for voice, including the expert-vetted crisis helpline support.
Special protections have been put in place for teenage users, with age-appropriate behaviors being directly integrated into the model. Parents can set up parental controls that allow them to block ChatGPT Voice on their child’s linked account, and parents can receive notifications in higher-risk situations indicating self-harm or suicidal intentions.
What Isn’t Included Yet
Currently, GPT-Live does not support voice in combination with video and screen sharing in ChatGPT. OpenAI confirmed it is working on implementing these capabilities, and in the meantime, the previous Standard and Advanced Voice Modes are still available for use when video or screen sharing is required.
According to the company, GPT-Live has been optimized for the most popular languages in ChatGPT, while performance may be lower for less widespread ones, resulting in a noticeable accent or reduced fluency.
Why This Matters for Search and Content
For everyone thinking about the visibility of their content in AI products, the new feature adds a new surface to consider. A question asked via voice can trigger a web search and result in the answer being returned right in the voice conversation, sometimes accompanied by a visual card on screen, without the need to even open a browser or click through multiple links.
What isn’t known yet is how sources will be cited for a spoken answer that requires web search. Typically, ChatGPT provides links to the sources next to its textual response, but OpenAI didn’t say anything about whether the spoken answer includes the sources aloud, visually, or doesn’t cite its sources at all. This is important for publishers and marketers because it would determine whether voice search within ChatGPT would still result in visits to the source, or the answer alone will suffice.
Frequently Asked Questions
Q1: What is GPT-Live?
GPT-Live is the new generation of voice models developed by OpenAI that powers ChatGPT Voice. The technology relies on full-duplex architecture allowing listening and speaking simultaneously and delegation of questions needing search and deeper reasoning to a frontier model like GPT-5.5.
Q2: When did GPT-Live launch?
OpenAI started rolling out GPT-Live on July 8, 2026, to all ChatGPT users around the world via iOS, Android, and web versions.
Q3: Which ChatGPT users get GPT-Live-1 versus GPT-Live-1 mini?
GPT-Live-1 is available for Go, Plus, and Pro subscribers by default. GPT-Live-1 mini is the default for Free users.
Q4: How does GPT-Live handle web searches during a voice conversation?
When the spoken question needs current information or deeper reasoning, GPT-Live sends it to GPT-5.5 behind the scenes and continues the conversation until the answer is ready and included in the voice exchange.
Q5: What are the visual cards in ChatGPT Voice?
ChatGPT Voice can now display visual cards for topics such as weather, stocks, sports scores, and maps while the user is talking, providing spoken answers with a visual counterpart.
Q6: Does GPT-Live support video or screen sharing?
Not yet. OpenAI said it is working on it, and the older Standard and Advanced Voice Modes are still available when video or screen sharing is required.
Q7: How did OpenAI evaluate GPT-Live’s performance?
OpenAI conducted head-to-head human evaluations of GPT-Live versus Advanced Voice Mode in five to ten minute conversations measuring overall preference, turn-taking, interruptions, and conversational flow, as well as benchmark tests like GPQA and BrowseComp.
Q8: What safety measures are built into GPT-Live?
OpenAI added audio-specific safety testing and evaluations for areas like self-harm, emotional reliance, and psychosis, as well as safeguards able to redirect the conversation, surface safety messaging or resources, or stop it in higher-risk situations.
Q9: How many people use ChatGPT Voice?
OpenAI said that more than 150 million people interact with ChatGPT each week via features like Voice and Dictation.
Q10: Does GPT-Live cite its sources when it searches the web?
OpenAI’s announcement did not specify whether spoken answers name sources aloud, display them visually, or omit them, leaving that detail unclear for now.
