OpenAI has launched GPT-Live-1, a new full-duplex voice model designed for ChatGPT, which significantly outperforms its predecessor, Advanced Voice Mode, across multiple areas. In particular, GPT-Live-1 excels in GPQA, testing expert-level scientific reasoning, and BrowseComp, which assesses the ability to locate difficult information via web search. This new model features a full-duplex architecture that allows it to listen and speak simultaneously, enhancing conversational flow and responsiveness. As part of the rollout, two versions of GPT-Live-1 will be made available, with API support expected to follow shortly.

OpenAI: OpenAI is an artificial intelligence research and deployment company focused on building advanced generative models and conversational systems. The company has launched GPT-Live-1 as its newest full-duplex voice model for ChatGPT, delivering improvements over Advanced Voice Mode on benchmarks for scientific reasoning, web navigation, and telecom support tasks. GPT-Live-1 is rolling out to ChatGPT users with API support planned shortly after.
GPT-Live-1: GPT-Live-1 is OpenAI’s latest full-duplex voice model designed for natural, real-time conversations in ChatGPT. It can listen and speak simultaneously, respond with conversational cues like “mhmm,” and delegate complex queries involving web search or deeper reasoning to frontier models. The model shows gains over prior Advanced Voice Mode on GPQA, BrowseComp, and internal telecom benchmarks.
GPT-Live-1 mini: GPT-Live-1 mini is the compact version of OpenAI’s new full-duplex voice model family for ChatGPT. It shares the same architecture enabling simultaneous listening and speaking along with delegation to more powerful models for demanding tasks. The mini variant is included in the rollout to all users in the coming days.

`json
{
“Benchmark Usage”: “Evaluations such as GPQA for expert scientific reasoning and BrowseComp for agentic web tasks are used to measure advanced AI capabilities beyond basic conversation.”,
“Voice Model Architecture”: “Full-duplex designs allow AI voice systems to process incoming speech while generating responses without waiting for users to finish speaking.”
}
`