Editorial illustration for OpenAI Releases GPT-Live Voice Models That Delegate Reasoning to GPT-5.5
OpenAI Launches GPT-Live Voice Models With GPT-5.5
OpenAI shipped a new voice model family Tuesday called GPT-Live, and it now runs the ChatGPT Voice experience for every user, everywhere, starting today. Two versions are rolling out first: GPT-Live-1 and a smaller GPT-Live-1 mini. Both replace Advanced Voice Mode, which OpenAI says lost head-to-head preference tests against the new models by a wide margin.
The pitch is a full-duplex model, one that listens and talks at the same time instead of waiting its turn. That's a break from how ChatGPT's voice mode has worked until now, where the system either chained together separate speech and text models or forced a stiff back-and-forth rhythm. GPT-Live handles the flow of conversation itself, including filler sounds and pauses, but it isn't meant to do the heavy lifting alone.
When a query needs a web search or real reasoning, GPT-Live hands it off to a separate frontier model running in the background, in this case GPT-5.5, and picks the answer back up once it's ready. An API version is planned, though video, screen sharing, and full language coverage aren't part of today's release.
Tapping the Voice button now uses the GPT-Live experience. Conversations support interruption, pausing, and requests to slow down. The model acknowledges you with cues like ‘mhmm’ or ‘got it.’ OpenAI remastered the nine distinct voices for GPT-Live.
Why this matters
Splitting voice and reasoning into separate models is OpenAI admitting the one-model-does-everything approach to real-time voice was never going to work well. Latency and reasoning depth pull in opposite directions, so handing the heavy lifting to GPT-5.5 while GPT-Live handles the conversational layer is a sensible division of labor. For developers building voice apps, this matters because it changes the integration story: you're no longer picking one model and hoping it's fast enough and smart enough.
You're building on a pipeline where the reasoning layer can be swapped out as OpenAI ships new frontier models, at least in theory. That's worth watching closely rather than taking at face value. The mini version also signals OpenAI wants this running on cheaper infrastructure, which matters for founders doing cost math on voice products at scale.
What we don't have yet is real-world latency data outside OpenAI's own evaluations, or clarity on how often the handoff to GPT-5.5 actually introduces lag users can hear. That's the number to track once developers get their hands on it.
Related Reading
Browse all OpenAI news →
LLMs & Generative AI ChatGPT's 'Nerdy' tweak rewards goblin metaphors in answers, study finds
LLMs & Generative AI Google tests visual 'magazine-style' UI for Gemini 3 Pro users
LLMs & Generative AI AI Engineers Face Rising Costs, Need New Strategies for Efficiency
Related Topic OpenAI hires 630 ex-Meta staff as ChatGPT memory may turn data into ads
Business & Startups
Related Topic Sen. Warren urges Trump admin to expand AI credit after OpenAI letter
Business & Startups
From Archives Bezos-backed startup raises USD 320 million to build AGI from gaming data
From Archives Anthropic Launches Claude Cowork AI Agent for Mobile and Web
From Archives Meta Tests GPT-5.5 With 'Watermelon' Model
From Archives GPT-4o, Gemini, and Claude Vision Can Reason Over Visual Details
Common Questions Answered
What is full-duplex technology in GPT-Live and how does it differ from Advanced Voice Mode?
Full-duplex technology allows GPT-Live to listen and talk simultaneously instead of waiting for its turn to speak, which is a significant departure from how Advanced Voice Mode operated. This capability enables more natural conversations where users can interrupt, pause, or request the model to slow down without waiting for the model to finish speaking first.
How does GPT-Live delegate reasoning tasks to GPT-5.5?
OpenAI split voice and reasoning into separate models, with GPT-Live handling the conversational layer and real-time voice interactions while delegating deeper reasoning tasks to GPT-5.5. This division of labor addresses the fundamental conflict between latency requirements and reasoning depth, allowing GPT-Live to maintain fast response times while GPT-5.5 handles complex analytical tasks in the background.
What conversational features are supported by GPT-Live voice models?
GPT-Live supports interruption, pausing, and requests to slow down during conversations, making interactions more natural and user-controlled. The model also acknowledges users with conversational cues like 'mhmm' or 'got it,' and OpenAI remastered nine distinct voices specifically for the GPT-Live experience to enhance voice quality.
How did GPT-Live perform compared to Advanced Voice Mode in testing?
GPT-Live defeated Advanced Voice Mode by a wide margin in head-to-head preference tests, demonstrating significant improvements in user experience and performance. Both GPT-Live-1 and GPT-Live-1 mini versions have now replaced Advanced Voice Mode across the ChatGPT Voice experience for all users globally.
What does the separation of voice and reasoning models mean for developers building voice applications?
The split between GPT-Live and GPT-5.5 changes the integration story for developers, as they are no longer selecting a single model and hoping it balances speed with capability. Instead, developers can leverage the specialized strengths of each model: GPT-Live for fast, responsive conversational interactions and GPT-5.5 for complex reasoning tasks.
Further Reading
- GPT-Realtime-2: A Voice Model with GPT-5-Class Reasoning - DataCamp
- OpenAI brings GPT-5-class reasoning to real-time voice - Venturebeat - VentureBeat
- Advancing voice intelligence with new models in the API - OpenAI
- GPT-Realtime-2: OpenAI's MOST Intelligent Voice Model Yet! - YouTube
- Reasoning models | OpenAI API - OpenAI Developers
