OpenAI has released GPT-Live-1.1, a new model for voice interaction in ChatGPT, alongside a lightweight version called GPT-Live-1.1 Quip. These models utilize full-duplex technology, allowing them to listen and respond simultaneously, which enables them to react to pauses, speech tempo, and interruptions in real time.
The primary GPT-Live-1.1 model is now the default for ChatGPT Plus subscribers and is integrated into the Codex and Dots services. Free-tier users are restricted to the lighter Quip version. Both models feature improved speech synthesis and pronunciation across multiple languages compared to their predecessors.
Internal testing by OpenAI revealed that while the new models perform similarly to previous versions in many areas, they show regressions in safety compliance. Specifically, performance declined regarding sensitive content, mental health, emotional dependency, and violence. The company plans to address these issues in future updates.
In tests regarding error reporting, the main GPT-Live-1.1 model was more likely to admit when a programming task was impossible, while the Quip version performed worse. When internet access was disabled, the lightweight model proved more reliable at notifying users of the limitation than the primary version. OpenAI confirmed that the models do not pose high-level threats in areas like biological weapons or cyberattacks.
