OpenAI Introduces 'GPT-Live,' A Model That Can Listen And Speak At The Same Time: Too Uncanny To Be Uncanny

The rapid development of conversational AI began from a quiet wage.

From the public release of ChatGPT by OpenAI in late 2022, that launch drew immediate attention to large language models (LLMs) and set off sustained competition among multiple technology companies to advance model capabilities, scale, and practical applications.

Over the following years, efforts expanded beyond text to include voice interfaces, with incremental improvements in speech recognition, response generation, and overall fluidity.

In 2024, OpenAI introduced an advanced voice mode alongside its GPT-4o model. One of the available voices, labeled Sky, prompted significant public reaction. Actress Scarlett Johansson, who had voiced an artificial intelligence character in the 2013 film Her, stated that the voice sounded eerily similar to her own.

She had previously declined an offer from OpenAI to provide her voice for the system. After widespread commentary and legal correspondence, OpenAI removed the Sky voice option.

Now, OpenAI announced 'GPT-Live,' a new family of voice models intended to power the voice experience inside ChatGPT, which probably could make Her a little bit too real, but without Johansson.

In the announcement, OpenAI said that:

The central technical change is a full-duplex architecture. Previous voice systems in ChatGPT largely operated on a turn-based or cascaded basis, in which the model typically waited for a clear end to user speech before generating a response.

GPT-Live instead processes incoming audio continuously while also producing output.

It's full-duplex architecture allows it to evaluate the audio stream many times per second and decides in real time whether it needs to speak, or continue listening, pause, interject briefly, or maybe hand off a task.

This design supports backchannel signals such as "mhmm" or "yeah" while a user is still speaking, more natural handling of interruptions, and the ability to remain silent during thinking pauses.

A second element separates real-time conversational handling from deeper computation.

When a query requires web search, extended reasoning, or other complex work, GPT-Live can delegate that portion to a GPT-5.5 at the time of launch—running in the background.

The voice model keeps the spoken exchange going without long silences and incorporates the results into the conversation once they are ready.

OpenAI states that this approach maintains conversational continuity even for questions that would previously have produced noticeable delays.

Live translation is listed among the capabilities enabled by the continuous audio processing.

The models also include remastered voice options with adjustments aimed at better performance in noisy environments and clearer focus on the primary speaker.

At launch, the system supports text, images, memory, and search within the same voice session, and visual elements such as cards for weather or other data can appear alongside spoken responses.

Early rollout observations from users have included comments on reduced latency and more responsive back-and-forth, alongside occasional notes about response length, personality consistency, or access to longer conversation history in the initial days.

OpenAI has described ongoing evaluation and refinement as the models reach wider availability.

The release introduces two variants at launch: GPT-Live-1 as the default for paid subscription tiers and GPT-Live-1 mini for free users.

These models are rolling out globally across the iOS and Android apps as well as the web version of ChatGPT, with API access planned for a later date.

The announcement positions GPT-Live as an incremental step that addresses specific constraints in earlier voice implementations through simultaneous audio handling and background delegation.