OpenAI Introduces GPT-Live-1.1
The new voice AI speaks better but adheres to rules worse
The key feature of the new models is full-duplex mode: they continuously listen to the interlocutor and respond while accounting for pauses, interruptions, and changes in speech tempo. The system independently determines whether to keep listening or transition to a response.
GPT-Live-1.1 has become the default model for paid plans Plus and above, while Quip is for other users. GPT-Live-1.1 replaced GPT-Live-1, and GPT-Live-1.1 Quip replaced GPT-Live-1 mini. Both models have improved speech sound and pronunciation in various languages; the main version also operates in Codex and Dots services.
In compliance checks, the models generally performed at the level of their predecessors but declined in four areas: sensitive content, mental health, emotional dependence on AI, and cruelty. The reasons for this are planned to be studied and addressed in future releases.
In tests for acknowledging its own mistakes, GPT-Live-1.1 more frequently reported the impossibility of a programming task, whereas Quip showed a slightly worse result. With internet access disabled, the junior version performed better, while the senior version performed worse. The tests were conducted without the protective mechanisms of publicly available versions.
