OpenAI Launches GPT-Live-1 to Collapse Voice AI Stack
The new native full-duplex API replaces cascaded speech systems with a single conversational layer capable of delegating complex reasoning to backend models.
OpenAI has released GPT-Live-1, a native full-duplex voice architecture available via API that allows AI to listen and speak simultaneously. The launch marks a fundamental shift in how voice agents operate by removing the latency and rigidity of traditional speech pipelines.
Unlike previous systems, GPT-Live-1 collapses the traditional "cascaded" voice stack—which required separate Speech-to-Text (STT), Large Language Model (LLM), and Text-to-Speech (TTS) components—into a single conversational frontline. This integration allows the model to handle mid-sentence interruptions and natural speech flow without needing separate system coordination. According to data from The New Stack, GPT-Live-1 performs 30 percentage points better than GPT-Realtime-2.1 on the Full Duplex Bench.
A Split-Brain Architecture
To maintain fluid conversation while performing complex tasks, GPT-Live-1 utilizes a delegation system. The model acts as the immediate interface, handling the nuances of human interaction, but can hand off heavy processing and tool calls to backend reasoning models, such as GPT-6 Astra or Luna, via an event-driven interface using a delegation_id. This allows the AI to acknowledge a speaker or fill pauses while a more powerful model works in the background to find an answer.
Industry Impact and Performance
This architectural shift significantly improves the human-like quality of AI interactions. The language-learning company Speak reported that GPT-Live-1 was nearly 80% less likely to interrupt users who paused to think compared to previous turn-based systems. By removing the need for developers to manually orchestrate three different models, OpenAI is reducing the engineering overhead required to build sophisticated voice agents.
Pricing and Deployment
OpenAI has priced the front-end voice layer at $0.05 per minute, which equates to approximately $3 per hour. However, developers must also account for the additional costs of any backend reasoning models used during the delegation process. While this reduces the local codebase for developers, it increases their structural dependency on OpenAI's proprietary platform.
What to Watch
As native multimodal models become the standard, the industry will likely move away from chained systems entirely. Future developments will likely focus on the efficiency of the delegation interface and whether other providers can match the performance gains seen on the Full Duplex Bench. It remains to be seen how widely the GPT-6 Astra and Luna backend integrations will be adopted across different enterprise use cases.