OpenAI Launches Full-Duplex Voice Models That Listen and Speak Simultaneously

OpenAI has released a suite of advanced voice models designed to replicate natural human conversation through simultaneous listening and speaking capabilities. The company introduced GPT-Live-1 and GPT-Realtime-2.1 across both consumer and developer platforms, addressing long-standing limitations in voice-based AI interactions. These full-duplex models reduce response latency by at least 25% and enable users to interrupt naturally without triggering system delays.

Consumer and Developer Release

OpenAI is rolling out GPT-Live-1 mini as the default Advanced Voice Mode replacement in ChatGPT for all users, with the larger GPT-Live-1 model reserved for paid subscription tiers. The architecture represents a fundamental shift from the previous three-step process—which separately chained speech-to-text, language model processing, and text-to-speech components. The unified design processes audio through a single model, preserving speech nuance while cutting computational overhead.

For developers, the company released GPT-Realtime-2.1 and GPT-Realtime-2.1-mini through its API. Both target low-latency voice and multimodal applications. The mini variant functions as a reasoning model for real-time voice, maintaining cost parity with its predecessor while supporting tool use and function calling—allowing the model to execute a task, verbalize progress, and deliver results without silent delays that typically confuse users.

Solving Real-World Voice Interaction Problems

The new models tackle critical friction points. Previously, voice agents would fall silent during tool calls, prompting users to assume connection failures and interrupt mid-operation. GPT-Realtime-2.1 counters this with configurable reasoning effort across five levels—minimal, low, medium, high, and xhigh—and spoken preambles. A model might now say "I'll check that order now" before processing, maintaining conversational coherence throughout multi-step requests. The 25% latency reduction stems from improved caching mechanisms across all Realtime voice models.

The larger GPT-Realtime-2.1 also improves alphanumeric recognition and refines silence and noise handling. Both models integrate access to newer text-based reasoning capabilities, including GPT-5.5 for search and agent-based tasks, while supporting visual response formatting for richer context delivery. OpenAI product lead Atty Eleti noted that internal testing has sustained 30- to 40-minute conversations during routine activities like walks, indicating the models' ability to maintain extended context without degradation.

Future of Voice as Primary Interface

OpenAI views voice as a foundational layer for complex computational work. The company has indicated that voice could eventually become the primary interface for managing lengthy, agentic tasks—the type of work users currently accomplish through code and text-based ChatGPT interactions. Reports suggest hardware deployment, potentially including AI-enabled earbuds, though the company has not disclosed specifics. The release underscores the growing adoption of voice features, with more than 150 million people already using ChatGPT's voice and dictation functions.