
September 2023 Releases
Earlier this week, OpenAI launched GPT-Live, a new generation of voice models, and described it as a “full-duplex” model.
This, in plain terms, means the model can listen and talk at the same time. In natural conversation, the model can acknowledge you mid-sentence, sit through a pause without jumping in, and pick back up when you finish a thought.
For anything that needs real reasoning or a web search, GPT-Live hands the task to GPT-5.5 in the background and folds the answer back into the conversation once it's ready. The conversation keeps moving while it handles deep reasoning.
OpenAI describes this move in their launch announcement video: "We are now pushing voice technology closer to how real people talk to each other. Where it can follow interruptions, pauses, corrections, and the natural flow of thinking this out loud."
The core idea: separate the model that has to be instantly responsive from the model that has to handle reasoning. This setup addresses the same tradeoff contact centers have lived with for years: respond fast, or respond correctly. With full duplex, the desire is for AI agents to get closer to doing both, while sounding more human.
This idea isn't new. Thinking Machines Lab proposed the same architecture as a research preview back in May. They proposed pairing a time-aware interaction model with a separate background model for deeper reasoning. GPT-Live's launch reads more like execution of that research.

In theory, GPT‑Live should also continuously process input while generating output. The model can therefore make several decisions per second, including “whether to speak, continue listening, pause, interrupt, or invoke a tool.”
We tested GPT-Live directly to see how the theory holds up in practice. In terms of the voice itself, it is natural sounding, allowing for genuine back-and-forth. However, several issues stood out enough to flag for anyone evaluating it for enterprise production use.
Action invocation: While GPT-Live is searching the web or executing a tool call, it still goes quiet. That's a gap from how OpenAI describes the delegation model, keeping the conversation moving while GPT-5.5 works in the background.
Latency: The pause during tool calls and search is noticeable, suggesting latency is still a real tradeoff even with full-duplex handling the rest of the conversation.
Conversational quality: GPT-Live reads very short feedback from callers as an interruption. Saying "ok" partway through one of its answers, meant as active listening rather than a new question, was enough to make it stop entirely.
This shift points to the research being done to help AI agents handle more complex tasks with conditional logic. As AI agents take on deeper reasoning, it reflects a broader industry desire: a more dynamic agent that can reach a goal on complex tasks without following a script, or sacrificing latency.
The gaps in GPT-Live's current performance, the silence during tool calls, the noticeable latency, the misread backchannels, are the visible seams in that transition. They're not signs the idea is wrong. They're signs the industry is still early in building agents that can reason and stay present in a conversation at the same time.
Earlier ChatGPT voice modes either chained separate speech-to-text, reasoning, and text-to-speech models together, or used a single model that still processed conversations in discrete turns. GPT-Live processes input and generates output continuously, and delegates deeper reasoning or search to GPT-5.5 in the background without pausing the conversation.
Enterprise voice AI agents face the same tradeoff between response speed and reasoning depth that consumer assistants do, often with higher stakes around compliance and accuracy. In theory, a split-model approach lets an agent stay responsive in the moment while still handling complex lookups or multi-step reasoning in the background.
Voice quality is natural and allows for genuine back-and-forth. In testing, a few gaps stood out: GPT-Live goes quiet while executing a tool call or web search, the pause during that delegation is noticeable, and short listener feedback like "ok" can get misread as a full interruption, cutting its response short.
Ready to see Regal in action?
Book a personalized demo.



